THE WORLD BANK
1 ECONOMICREVIEW I
EDITOR
Jaime de Melo, University of Geneva
EDITORIAL BOARD
Kaushik Basu, CornellUniversity,USA Paul Gertler, WbrldBank
Jean-Marie Baland, Universityof Namur, Indermit GiU,World Bank
Belgium Jan Willem Gunning, Free University,
Chong-En Bai, TsinghuaUniversity, The Netherlandr
China Jeffrey Hammer, World Bank
Timothy Besley, London School of Graciela Kaminsky, George Washington
Economics,UK University, USA
Fran~oisBourguignon, World Bank Ravi Kanbur, CornellUniversity, USA
Kenneth Chomitz, World Bank Peter Lanjouw, World Bank
Maureen Cropper, UniversityofMaryland, Thierry Magnac, Universitide TouloureI,
USA France
Jishnu Das, World Bank Juan-Pablo Nicolini, UniversidadTorcuato di
Klaus Deininger, WorldBank Tella, Argentina
Asli Demigur5-Kunt, World Bank
Boris Pleskovic, World Bank
Stefan Dercon, Oxford University, UK
Martin Ravallion,World Bank
Ishac Diwan, World Bank
Ritva Reinikka, World Bank
Augustin Kwasi Fosu, UNEconomic
Elisabeth Sadoulet, University of Cal@nia,
CommissionforAfiica (ECA),Ethiopia
Berkeley, USA
Alan Harold Gelb, World Bank
Mark Gersovitz,Johns Hopkins Joseph Stiglitz, Columbia University, USA
University, USA L. Alan winters, World Bank
The World Bank Economic Review is a professional journal for the dissemination of World
Bank-sponsored and other research that may inform policy analysis and choice. It is directed
to an international readership among economists and social scientists in government, business,
international agencies, universities, and development research institutions. The Reviewseeks to
provide the most current and best research in the field of quantitative development policy
analysis, emphasizing policy relevance and operational aspects of economics, rather than
primarily theoretical and methodological issues. It is intended for readers familiar with
economic theory and analysis but not necessarily proficient in advanced mathematical or
econometric techniques. Articles illustrate how professional research can shed light on policy
choices. Consistency with World Bank policy plays no role in the selection of articles.
Articles are drawn from work conducted by World Bank staff and consultants and by outside
researchers.Non-Bank contributors are encouraged to submit their work.
Before being accepted for publication, articles are reviewed by three referees-one from the World
Bank and two from outside the institution. Articles must also be endorsed by two members of the
Editorial Board before final acceptance.
For more information,please visit the Web sites of the Econorriic Reviewat Oxford UniversityPress
at www.wber.oxfordjournals.org and at the World Bank at www.worldbank.org/research/journals.
Instructionsfor authorswishingto submitarticlesare availableonline atwww.wber.oxfordjournals.org.
Plcasc direct alleditorialcorrespondenceto the Editor at wbeKLorldbank.org.
THE WORLD BANK ECONOMIC REVIEW
Volume 20 2006 Number 1
Making Conditional Cash Transfer Programs More Efficient:
Designing for Maximum Effect of the Conditionality 1
Alain de Janvry and Elisabeth Sadoulet
Child Labor and School Achievement in Latin America 31
Victoria Gunnarsson, Peter F. Orazem, and Mario A. Sanchez
The Long-Run Economic Costs of AIDS: A Model with an
Application to South Mica 55
CIive Bell, Shantayanan Devarajan, and Hans Gersbach
Robust Multidimensional Spatial Poverty Comparisons in
Ghana, Madagascar, and Uganda
Jean-Yves Duclos, David Sahn, and Stephen D. Younger
The Impact of Regulatory Governance and Privatization on
Electricity Industry Generation Capacity in Developing Economies 115
John Cubbin and Jon Stern
An Empirical Analysis of State and Private-Sector Provision of
Water Services in M i c a 143
Colin Kirkpatrick, David Parker, and Yin-Fang Zhang
SUBSCRIPTIONS: A subscription to The World Bank Economic Review (ISSN 0258-6770) comprises
3 issues. Prices include postage; for subscribers outside the Americas, issues are sent air freight.
AnnualSubscription Rate (Volume20,3Issues,2006):Academic librariedrint editionand site-wideonline
access:US$137/E87/131, Print editiononly:US$131K83/125, Site-wideonlineaccessonly:US$123/08/
117; Corporate-Print edition and site-wideonlineaccess: US$205/E130/195, Print editiononly:US$196/
.&124/186, Site-wide online access only:US$185/E117/176; Personat--Printedition and individual online
access: US$49/E34/51. Pleasenote:ESterlingratesapplyinthe UK, 6 in Europe,and US$elsewhere.There
may be other subscription rates available; for a completelisting, please visit &.wber.oxfordjournals.org/
subscriptions.Readers with mailingah'resses in non-OECDcountriesand in socialisteconomies in transition are
eligibleto receive complimentsy subscriptionson reyuest by writing to the UKaddressbelow.
Full prepayment in the correct currency is required for all orders. Orders are regarded as firm,and
payments are not refundable. Subscriptionsare accepted and entered on a complete volume basis. Claims
cannot be considered more than four months after publication or date of order, whichever is later. All
subscriptionsin Canada are subject to GST. Subscriptions in the EU may be subject to European VAT.
If registered, please supply details to avoid unnecessary charges. For subscriptions that include online
versions, a proportion of the subscription price may be subject to UK VAT. ~ersonalrates are applicable
onlywhen a subscription is for individual use and are not availableif deliveryis made to a corporate address.
BACK ISSUES: The current year and two previous years' issues are available from Oxford UniversityPress.
Previous volumes can be obtained from the Periodicals Service Company,11Main Street, Germantown,
NY 12526, USA. E-mail: psc@periodicals.com.Tel: (518) 537-4700. Fax: (518) 537-5899.
CONTACTINFORMATION: JournalsCustomerServiceDepartment, OxfordUniversityPress,Great Clarendon
Street,Oxford OX2 6DP,UK. E-mail:jnls.cust.serv@oxfordjoumals.org.Tel: +44(0)1865353907. Fax: + 44
(0)1865353485. In theAmericas,please contact: Journals Customer Service Department, Oxford University
Press, 2001 Evans Road, Cary, NC 27513, USA. E-mail: jnlorders@oxfordjournals.org.Tel: (800) 852-
7323 (toll-free in USAICanada) or (919) 677-0977. Fax: (919) 677-1714. In Japan, please contact: Journals
Customer Service Department, Oxford UniversityPress,1-1-17-SF, Mukogaoka,Bunkyo-la, Tokyo,113-
0023,Japan. E-mail: okudaoup@po.iijnet.or.jp.Tel: (03) 3813 1461. Fax: (03) 3818 1522.
POSTAL INFORMATION: The World Bank Economic Review (ISSN 0258-6770) is published by Oxford
University Press for the International Bank for Reconstruction and DevelopmendT~EWORLD BANK.
Send address changes to The World Bank Economic Review, Journals Customer Service Department,
Oxford University Press, 2001 Evans Road, Cary, NC 27513-2009. Communications regarding original
articles and editorial management should be addressed to The Editor, The World Bank Economic Review,
The World Bank, 3, Chemin Louis Dunant, CP66 1211 Geneva 20, Switzerland.
DIGITAL OBJECT IDENTIFIERS: For information on dois and to resolve them, please visit www.doi.org.
PERMISSIONS: For information on how to request permissions to reproduce articles or information from
this journal, please visit www.oxfordjournals.org/jnls/permissions.
ADVERTISING: Inquiries about advertising should be sent to Helen Pearson, Oxford Journals Advertising,
PO Box 347, Abingdon OX14 lGJ, UK. E-mail: helen@oxfordads.com.Tel: +44 (0)1235201904. Fax:
+44 (0)8704 296864.
DISCLAIMER: Statements of fact and opinion in the articles in The World Bank Economic Review are
those of the respectiveauthors and contributors and not of the International Bank for Reconstruction and
develop mend^^^ WORLD BANK or Oxford University Press. Neither Oxford University Press nor the
International Bank for Reconstruction and DevelopmendTHE WORLD BANK make any representation,
express or implied, in respect of the accuracy of the material in this journal and cannot accept any legal
responsibilityor liabilityfor any errors or omissions that may be made. The reader should make her or his
own evaluation as to the appropriateness or otherwise of any experimentaltechnique described.
PAPER USED: The World Bank Economic Review is printed on acid-free paper that meets the minimum
requirements of ANSI Standard 239.48-1984 (Permanence of Paper).
INDEXING AND ABSTRACTING: The World Bank Economic Review is indexed andlor abstracted by CAB
Abstracts, Current ContentdSocialand Behavioral Sciences, Journal of Economic LiterahrdEconLit, PAlS
IntemationaL RePEc (Research in Economic Papers), and Social Sewices Citation Index.
COPYRIGHT O The International Bank for Reconstruction and Development/TH~WORLD BANK 2006
All rights reserved; no part of this publication may be reproduced, stored in a retrieval system, or
transmitted in any form or by any means, electronic, mechanical, photocopying, recording, or otherwise
without prior written permission of the publisher or a license rkstricted copying issued in the
UK by the Copyright Licensing Agency Ltd, 90 Tottenham Court Road, London W1P 9HE, or in the
USA by the Copyright Clearance Center, 222 Rosewood Drive, Danvers, MA 01923.
Making Conditional Cash Transfer Programs More
Efficient: Designing for Maximum Effect of the
Conditionality
Alain de Janvryand Elisabeth Sadoulet
Conditional cash transfer programs are now used extensivelyto encourage poor parents
to increase investments in their children's human capital. These programs can be large
and expensive, motivating a quest for greater efficiency through increased impact of the
programs' imposed conditions on human capital formation. This requires designing the
programs' targeting and calibration rules specifically to achieve this result. Using data
from the Progresa randomized experiment in Mexico, this article shows that large
efficiency gains can be achieved by taking into account how much the probability of
a child's enrollment is affected by a conditional transfer. Rules for targeting and
calibration can be made easy to implement by selecting indicators that are simple,
observable, and verifiable and that cannot be manipulated by beneficiaries.The Mexico
case shows that these efficiency gains can be achieved without increasing inequality
among poor households.
Conditional cash transfer programs targeted to poor households have become
widely used, in particular to induce beneficiary households to invest in their
children's human capital. The approach presumes that the supply side of social
services for education and health is in place and that stimulating demand
through income effects is insufficient to induce major changes in human capital
investment (Bourguignon, Ferreira, and Leite 2002). Instead, a condition that
transforms the income effect into a price effect needs to be attached to the cash
transfer. In this case, receiving the transfer requires meeting school attendance
and health practice requirements.
This approach has been hailed as a major innovation in organizing poverty-
reduction programs. Well-known programs that follow this approach include
Progresa (nowcalled Oportunidades) in Mexico, Bolsa Escola and Bolsa Familia
in Brazil, Red de Protecci6n Social in Nicaragua, Prograrna de Asistencia
Alain de Janvry and Elisabeth Sadoulet are professors in the Department of Agricultural and Resource
Economics at the University of California at Berkeley. Their e-mail addresses are alain@are.berkeley.edu
and sadoulet@are.berkeley.edu. They are grateful to Harold Alderman, Jean-Louis Arcand, Gabriel
Demombynes, Frederico Finan, Patrick Guillaumont, Sylviane Jeanneney-Guillaumont, John Maluccio,
Martin Ravallion, Kalanidhi Subbarao, and Miguel Szekely for their helpful comments.
THE WORLD BANK ECONOMIC REVIEW, VOL.20, NO. 1, pp. 1-29 doi:lO.l093/wber/lhj002
Advance Access publication February 1,2006
OThe Author 2006. Published by Oxford University Press on behalf of the International
Bankfor Reconstructionand Development/ nn.WORD BANK. All rights reserved. For permissions,
please e-mail: journals.permissions@oxfordjournals.org.
2 THE W O R L D BANK ECONOMIC REVIEW, VOL. 20, N O . I
Familiar in Honduras, the Program of Advancementthrough Health and Educa-
tion in Jamaica, Food-for-Education in Bangladesh, and SubsidioUnico Familiar
in Chile (Ravallion and Wodon 2000; Skoufias 2000; Morley and Coady 2003;
Rawlingsand Rubio 2005).Some of these programs have become very large and
expensive. In 2004 Oportunidades served 4 million families at a cost of $2.2
billion. In 2001 Bolsa Escola covered 4.8 million families at a cost of $700
million. While few programs have been rigorously evaluated, Progresa was
found to have a positive impact on education (Schultz2004) and health (Gertler
2004). However, almost no analysis has been conducted on the effectiveness of
alternative program designs in achieving these results, despite the large sums
spent to obtain them. This issue is addressed here by analyzing whether better
targeting of qualifying poor households and better calibration of the levels of
cash transfers could help raise program efficiency.
Conditional cash transfer programs have a dual objective: immediate poverty
reduction through transfers and long-term poverty reduction through investment
in human capital. Efficiently meeting the first objective requires the transfers to
be accurately targeted to poor households-a difficult task not addressed here
(van de Walle 1998; Alderman 2001, 2002; Ravallion 2003). Meeting the
second objective requires accurately selecting among poor households to mini-
mize the efficiency leakages from payments to children already highly likely to
attend school without a transfer (as opposed to children who will attend school
only because of the transfer) and offering a level of transfer that is sufficient to
meet the opportunity cost of the change in behavior, thus securinga high uptake
while minimizing project costs. The specific concerns addressed here are the
definition of targeting and calibration rules that can be easily implemented and
that are cost-effective and transparent, and the potential tradeoffs between
efficiencygains through implementation of these rules and higher inequality in
transfers among poor households.
Progresa is used to explore these alternative program designs. Efficiency gains
of 29-44 percent over the current program are found to be possible without
increasing inequality among poor households.
Progresa is a conditional cash transfer program for human capital formation
targeted to poor rural children. It consists of three closely related components to
address education, health, and nutrition issues: (a)a monetary transfer to each
child under age 18 who regularly attends school between the third year of
primary school and the third year of secondary school and who regularly visits
a health center, (b)basic healthcare for all family members, and (c)nutritional
supplements for children and women in need.
Progresa was introduced in 1997 and by 2000 fully covered marginal rural
municipalities, reaching 2.6 million families. The budget for 2000 was $950
million44 percent of it for education transfers, which benefited approximately
de Janvry and Sadoulet 3
1.6 million children in primary school and 800,000 in secondary school (Coady
2000).
The transfers that Progresa families receive significantly increase their
income-by an average of 22 percent. Progresa has explicitly targeted poor
households in marginal rural areas of Mexico. The purpose here is not to
question this targeting, which corresponds to Progresa's objectiveof transferring
resources to poor families. The purpose is to explore whether, for a given budget
constraint, targeting and calibrating transfers to poor households can more
efficiently increase school participation. Consequently, Progresa's education
component is used as a laboratory to explore alternative targeting and calibra-
tion rules. The idea is to derive lessonsfrom this richly informed experiment that
can be applied to Progresa and to other conditional cash transfer programs
where severely limited budgets make targeting critical.
To measure its impact, Progresa selected a sample of 506 marginal commu-
nities containing 24,000 householdsand 17,000 children eligiblefor transfers and
surveyed them a year before the program started and, subsequently, every six
months over three years. Information was collected on individual,household, and
community characteristics.The sample design consists of the random selection of
320 treatment communities and 186 control communities. The analysis is
restricted to children in school in October 1997. Twelve percent of eligible
children had dropped out of school by that time, some several years earlier, and
while the program helped bring some of them back to school, this one-time effect
at the onset of the program is not the focus of the analysis.The sample was further
restricted to the 2,242 poor children who graduated from primary school in the
summer of 1998 and faced the decision of whether to continue to secondary
school. These data are used to estimate a model of the school enrollment decision
that captures, in particular, the impact of Progresa transfers, paying particular
attention to heterogeneity of conditions among chldren. Alternative targeting and
transfer programs are then simulated and their efficiency is compared.
Focusing on Entry into Secondary School
Asimple analysisof the overall Progresa budgetsuggeststhat an efficientprogram
for schoolenrollmentshouldfocusstrictly on the transition from primaryschool to
secondary school, a point already suggested in the International Food Policy
ResearchInstitute evaluations (Skoufias2000; Coady 2000; Schultz 2004).
The conditional transfer offered to each child is calculated according to the
program's rules.' The program has a schedule of education transfers that
increase as children progress to higher grades and that are higher for girls than
for boys in secondary school (table 1).There is, however, a maximum total
1. The 'conditional transfer' isthe exogenous amount that achild would receivefrom Progresaif he or
she was in a treatment community and attended school. It depends on the gender and grade of the child
and the household's demographics. At the household level the conditional transfer is the total amount that
the household would receive if it was in a program community and complied with all Progresa rules.
TABLE 1. Budget for Educational Transfers: Progresa in Sample Villages, 1998
Annual Budget for Enrolled
Children
School Years Number of Schedule of Annual Continuation Share
that Children Eligible Conditional Transfer Rate U.S. of Total
Could Attend Childrena (us$) (percent) ~ o l l a r s ~ (percent)
Primary 3 1,909 70 98.2 114,229 11.8
Primary 4 1,811 80 97.8 120,260 12.4
Primary 5 1,613 100 97.1 135,626 14.0
Primary 6 1,476 135 97.4 166,035 17.2
Secondary1 1,416 20O/21Oc 76.7 189,602 19.6
Secondary 2 752 210/235' 96.1 134,884 14.0
Secondary 3 551 220/255' 96.7 106,028 11.0
Total 9,528 966,664 100
"Only children enrolled in school in 1997.
baking into account the cap on total household conditional transfers. With 10 monthly conditional transfers per school year and an exchange rate in
October 1998 of 10 pesos per US$, all transfers can be read as either pesos per month or US$ per year.
'Conditional transfers for boys/girls.
Source:Authors'analysis based on the 1998 Progresa survey.
de Janvry and Sadoulet 5
conditional transfer that each household can receive, set at $625 a year in 1998
(including $100 for n~trition).~In the sample, 13.4 percent of eligible children
are subject to the household transfer cap. Using the proportionality rule that
Progresa applies, the conditional transfer corresponding to each child is calcu-
lated by scaling down all the education transfers in a household subject to the
cap by the same f a ~ t o rAmong the children graduating from primary school,
. ~
28 percent are subject to the cap, and the conditional transfers vary from $100 to
the full $200/$210, with an average of $169. The budget for education transfers
in the sample treatment communities is calculated using these conditional trans-
fers and the enrollment status of each child (table 1).The budget would be
17 percent higher, with no cap on total household conditional transfer. Taking
into account these caps, transfers to primary school children account for 55.4
percent of the total budget for education transfers and the first year of secondary
school for almost 20 percent.
Other studies show that Progresa's conditional transfers increase continuation
rates at all grades (Behrman, Sengupta, and Todd 2001; de Janvry, Finan, and
Sadoulet 2001; Schultz 2004).However, school continuation rates were already
very high in both primary and secondary school before the Progresa intervention
(figure 1). The increase in continuation rates that result from the conditional
transfers is only around 1 percentage point in primary school and half a percen-
tage point in the second and third years of secondary school. This suggests that,
from an efficiency standpoint, Progresa is unnecessarily expensive for primary
school, with 96 school-attending children paid for each child that is retained in
school by the conditional transfer incentive, for an implied effective cost per
additional child attending primary school of $9,600. Assistingthe 3-4 percent of
children who drop out of school at each grade level would require a very
different program. Eliminating all transfers to primary school students would
have saved 55.4 percent of the education transfer budget, or more than $230
million of the $950 million budget in 2000.
Thecriticalproblemintermsof educationachievementoccursat entryinto lower
secondaryschool. Thus the analysis here continuesonly for secondary school.
The Efficiency of Progresa's Education Transfers
There are two sources of inefficiencyin a conditional cash transfer program that
need to be optimally reduced. The first is paying people for what they were
already going to do. With Progresa, this is obvious in primary school. But the
problem also arises in secondary school: 64 percent of the poor children who
graduate from primary school would enter secondary school without a transfer.
Reducing this efficiency leakage requires being able to predict who will not
2. This cap was introduced so that the program does not induce a fertility response.
3. For households subject to the cap all conditional transfers are scaled down by a common factor so
that they add up to the cap. This prevents the households from keeping a child out of school without
penalty.
FIGURE 1. SchoolContinuationRatesof Poor ChildreninSampleCommunities
Continuation rate (%)
Progresa communities
Control communities ',
I I
PROGRESA INTERVENTION
1
Primary school Lower secondary / ~ ~ ~ e ; s e c o n d a r ~
40 , I ! I I I I ! I 1 I
Entering grade
Source: Authors' analysis based on the 1998 Progresa survey.
continue in school, using a model for the probability that a given type of child
will enroll in school. Because such predictions are inevitably noisy, there is no
way to completely avoid this ineffi~ienc~.~The challenge, however, is to reduce
it by not targeting children most likely to attend school anyway.
The second source of inefficiencycomes from offeringtransfers that are either
too high or too low relative to the minimum amount needed to induce the
conditional action. As shown later, the simple difference estimation of the
impact of Progresa indicates that the program raised the enrollment rate from
63.6 to 76.6 percent. The conditional transfers offered were thus sufficient for
the 13 percent of children in the sample who were attracted to enroll in second-
ary school and would not have done so otherwise. Could a smaller conditional
transfer have had the same effect? For the 23.4 percent that did not take the
conditional transfer, would a higher transfer have induced them to accept the
offer?If so and if these children can be identified,should the conditional transfer
offered to them be increased?
4. This inefficiency concept is analogous to the issue of fungibility with inframarginal transfers,
whereby beneficiaries substitute other commodities for those subsidized by the program, meaning that
the program has no real effect on total consumption of the targeted commodity.
de Janvry and Sadoulet 7
If there were a clear opportunity cost to children's time in school, the subsidy
could be calibrated to match it. For Progresa the transfer is 40 percent of what
children of the same age earn when they work. But the opportunity cost of
children's time at school is not easy to calculate. Less than 30 percent of the
children who drop out at the end of primary school work during the subsequent
18 months (45 percent of boysand 10 percent of girls), increasing to 35 percent
(55percent of boys and 12 percent of girls)the following year. Lack of money or
need to work is the most common reason given for not continuing school (57
percent), but other important reasons are that the child does not like school or
does not learn (23percent) and that the school is too far away (13 percent).
The Progresa randomized experiment allows the estimation of the children's
response function, which is necessary to design the transfer schedule that max-
imizes return to the program. Since no experiments were conducted to observe
the response to different levels of conditional transfers, the cap on total condi-
tional transfers to a household is used to infer the marginal response to varying
conditional transfer amounts.
Dealing with these two sources of inefficiency requires an accurate predictive
model of the probability of attending school as a function of the characteristics
of the child, the household, and the community and of the amount of the
conditional transfer offered. The analysis here concentrates on entry into sec-
ondary school since that is where the conditional transfer can induce an impor-
tant behavior change that results in efficiency gains.
The conditional transfers offered to children in second and third years of
secondary school are not questioned for two reasons. First, these conditional
transfers are part of the expected benefits of entering secondary school, and the
measured impact of Progresa thus includes their effect. Second, while very high
continuation rates are observed in secondary school, they are for a selected
group of children who voluntarily entered secondary school without any subsidy
in 1998, before the program was in place. Other children who are induced to
enter secondary school with a conditional transfer are very unlikely to continue
at the same rate if the subsidies were discontinued. No experimental design
allows this particular continuation rate to be studied here because Progresa
always supported the first three years of secondary school. The safe bet is that
whatever transfer is provided in the first year needs to be provided in all three
years of secondary school (as is currently the case), and while in 1998 many
fewer Progresa children were in the second and third years of secondary school
than in the first year (becauseit was the first year of the program),these numbers
should even out three years after program implementation. Thus the results of
the analysis of the first year are applied to all three years of secondary school.
Let P(X,Y) denote the probability that a child with characteristics X and
eligible for a conditional transfer T will enroll in school. Eligibility is denoted
by the index function I E [0,11.Children's characteristics are distributed accord-
ing to the density function f(X).
The allocation problem consists in choosing the eligibility status I(X)and, if
eligible, the conditional transfer T(X)offered to a child with characteristics X, to
maximize the gain in enrollment over the population:
max JiP(x,T)- P(X,O)]I~(X)d~
I(X)>T(X)
subjectto a budget constraint:
where B is the budget available for the program. The first order condition for the
optimal conditional transfer is that, for any eligible child (I= I),
where PT= dP/aT and I is the Lagrange multiplier associated with the budget
constraint. This relationship states that the ratio of cost (PTT+ P)dT
to enrollment benefit (PTdT) of a marginal increase in the conditional
transfer (dT) is equal for all children. Hence, the cost of the marginal child
brought to school is equal across children types X. Note that the cost has two
terms. The first term, PTTdT,is the transfer cost to the PTdTmarginal children
brought to school by the increase in conditional transfer. The second term, PdT,
is the cost of giving the increasein transfer, dT, to all P children of the same type
X, even though they enrolled in school with the initial transfer T. This is the
marginal equivalent of the decomposition of the cost of transfers:
where the first term is the cost of the transfersto the children brought to school by
the conditional transfer and the second term the cost of transfers to the children
with similar observable characteristics who would have gone to school anyway.
Given the optimal conditional transfer amount conditional on eligibility, the
optimal eligibility rule is defined by:
Z = 1 if[P(X,T) - P(X,0)]T IP(X,T)T 2 0
+
(5) 1 = 0 otherwise.
The optimal allocation of a budget B is thus the solution to the system of
equations (2), (3),and (5).
In the particular case of a linear probability model, which is used in
the empirical work here, the conditional expectation of the enrollment prob-
ability is:
de Janvry and Sadoulet 9
where 601 +X P is the total impact of T and XS, which includes a constant
term, is the marginal impact of T. The presence of an intercept Sol is due to the
fact that only conditional transfers of $100-$210 are observed, meaning that the
linearity of the conditional transfer effect cannot be extended below that range
to a 0 conditional transfer.
The optimal conditional transfer and eligibility criteria defined in equations
(3)and (5)are:
where 1is the solution to the budget constraint in equation (2).The expression
shows that both eligibility and the optimal conditional transfer for any given
child are a function of the ratio
The first term in the numerator is the expected probability that children with
characteristics X would attend school even without a conditional transfer, and
the denominator is the marginal effect of the conditional transfer on the
expected enrollment probability. Children will thus be eligiblefor and be offered
high conditional transfers when they have a low initial probability of enrollment
or a high enrollment response to a conditional transfer. This optimal conditional
transfer is a function of all the characteristics X that predict enrollment, albeit in
a very nonlinear form.
Whether any program could use such a complex formula to compute condi-
tional transfers is questionable. But it is a useful benchmark because it gives the
maximum efficiency that could be reached with the observables X, and it will
thus be computed in the empirical analysis that follows. Next, however, the
optimal program is constrained to be linear and to use a small number of
observable characteristics.
An Implementable Conditional Cash Transfer Program
To be useful for program implementation, eligibility rules need to be simple and
transparent. Indicators used to determine eligibility, and the level of conditional
transfers must be few, easily observable and verifiable, and nonmanipulatable by
households. Simplicity and transparency are also important to ensure the political
acceptabilityof a subsidyprogram (Schady2002).Progresa usesgrade and gender to
set the schedule of conditional transfers (table1).The objective is thus to simplify
the formula in equation (7) established for the optimal conditional cash transfer
program to a linear index on the basis of a few characteristicsZ of the children.
1 0 T H E W O R L D B A N K E C O N O M I C REVIEW, VOL. 20, N O . I
The allocation problem lies in choosing the eligibility status and, if eligible,
the conditional transfer to offer to each child to maximize the gain in enroll-
ment over the population [equation (I)], subject to the budget constraint
[equation (2)],and using simple linear formulas for eligibility and conditional
transfer:
T = Za and I =IIZy 2 ymi,]
where Z is a subset of characteristics of the children and a, y, and ymi, are
parameters to be determined. As in the model above, optimal eligibility is
defined by the sign of the optimal conditional transfer value:
The parameters a are a solution to the maximization of a quadratic function:
Cm
(10) 1
max d i a - 1 (Ph b0
+ +miZia)Zia
a ieE
where E is the set of eligible children, mi = X,S is the marginal effect of the
conditional transfer on child i's school enrollment, Poi= Xipis child i's enroll-
ment probability without the transfer, and A is the Lagrange multiplier on the
budget constraint. The conditional transfer formula in equation (9) is thus a
simple linear combination of a few observed characteristics Z. It is similar to the
scoring system used in many welfare programs, whereby characteristics Z com-
mand scores a that add up to an aggregate score Za. In this case Za determines
not only eligibility but also the conditional transfer amount.
Important empirical questions are whether the use of this simple scoring
schedule is close enough to the optimal conditional cash transfer schedule and
what type I (exclusion)and type I1(inclusion)errors are made in this implemen-
tation. These questions are addressed after the schedules are established.
In this section a predictive model of entry into secondary school is built.
Although a probit and a logit perform better at high and low probabilities, a
linear model is used here to avoid imposing heterogeneity on the impact of the
conditional transfer through the functional form, since this will be an impor-
tant determinant of the targeting rule.' The sample of children finishing pri-
mary school and eligible for a Progresa transfer (defined as poor according to
5. In the simulation exercises that follow, the problem of predicted negative probability is never
encountered (the majority of children have predicted probabilities above 0.4), but some predictions are
above 1, even without conditional transfer and more when applying conditional transfers. For simulation
purposes these will be set equal to 1.
de Janvry and Sadoulet 11
the Progresa welfare index) in both the control and treatment communities is
used, and randomization in the selection of communities ensures that being in a
treated community is orthogonal to the children's ~haracteristics.~The average
treatment effect can thus be obtained simply by comparing the average enroll-
ment rates of children in the two types of communities. However, the actual
amount of conditional transfer offered to a child is not orthogonal to its
characteristics because being subject to the cap rule and the corresponding
household scaling factor are both a function of the children's age structure,
which is likely correlated with household preferences that influence schooling
decisions. The impact of the continuous treatment effect is thus estimated,
controlling for the conditional transfer level.
The empirical equivalent to equation (6)is written as:
or
(11) Si = 601, +61uTi+PoTi+Xip+ui,with control variables Xi
where S, is a binary variable indicating the enrollment status of child i, I, is a
dummy variable that indicates whether child i lives in a treatment community, T ,
is the conditional transfer that child i is eligible for under the program, and the
control var~ablesX, are child, household, and community characteristics.
The program increases the probability that qualifying poor children will
continue to secondary school by 13 percentage points [table 2, column (I)].As
expected this is slightly higher than the estimated 8-9 percentage point increase
in enrollment probability conditional on completed primary school (i.e., includ-
ing children who had dropped out of school before the program started)
obtained in other studies (e.g., Schultz 2004).
The value of the conditional transfer (variable I,T,), which varies across
children because of the cap on household transfer that affects 26 percent of the
qualifying children, is used in column (2). Kesults show that the marginal
effect of the conditional transfer is high (1.42 percentage points per $10).
Note that the imposed linear form gives a meaningful positive effect only for
conditional transfers above $100, which is not really restrictive because cur-
rent conditional transfers are much higher. Adding a large number of child,
household, and community controls in column (3) indicates that the main
correlates of a child's secondary school enrollment are age of the child (nega-
tive), mother's literacy and the household's maximum education level (posi-
tive), the number of agricultural workers and self-employed workers in the
household (negative), total expenditure (positive), and distance to school
(negative). State effects are also important. Both models predict that the
current $200 conditional transfer increases the probability of enrollment by
6. The quality of the randomization is verified and documented in Behrman and Todd (1999).
TAB LE 2. Linear Probability Model of Enrollment
Homogenous Impact Heterogeneous Impact
Variable Mean (1) (2) (3) (4) (5)
Treatment community
(dummy variable)
Conditional transfer"treatment
($100 per year)
Conditional transfer"treatment"
male
Conditional transfer"treatment*
(age- 12)
Conditional transfcr".treatmetIt"
father indigenous
Conditional transfer+treatment"
no secondary school in village
Child and household
characteristics
Conditional transfer
($100 per year)
Male
Age
Father is indigenous
Birth order
Head is male
Has no father
Father is literate
Father's education
(Continued)
TABL 2. Continued
E
Homogenous lmpact Heterogeneous Impact
Variable Mean (1) (2) (3)
Guerrero
Michoacan
Puebla
Queretaro
San Luis Potosi
Veracruz
Constant
Number of observations
R~
--- -
pp
Note: Numbers in ~arenthesesare standard errors.
"Significantat the 10percent level;
*'significant at the 5 percent level;
"""significant at the 1percent level.
Source:Authors' analysis based on the 1998 Progresa survey.
de lanvry and Sadoulet 15
the same 14 percentage points (table 3),which confirms that th~econtrols are
orthogonal to the treatment.7
Columns (4) and (5) explore heterogeneity of impact across categories of
children, with and without controls, focusing on the aspects of heterogeneity
that may be usable for targeting purposes. They are the child's age,' father's
ethnic it^, and whether there is a secondary school in the community. Progresa
recognizesgender differences, which are not found to be important in explaining
differential impacts of transfers on the decision to continue into secondary
school.' The results show that age, ethnicity, and presence of a school in the
community all have largeimpacts on enrollment, both directly as controls and in
affecting the impact of the conditional transfer. The results in column (5)are
used as the predictive model to evaluate the impact of targeting on enrollment.
Heterogeneity implies large differences in the impact of a conditional transfer
on enrollment acrosscategories of children (table 3).For a 12-year-old boy with
a nonindigenous father and a school in the community, the $200 conditional
transfer increases the probability of enrollment by only 3-4 percentage points. If
he is two years behind normal progress, the conditional transfer increases the
probability of enrollment by 10-12 percentage points. For a 12-year-old boy
with an indigenous father or no secondary school in the community, the condi-
tional transfer increasesthe probability of enrollment by 9-11percentagepoints.
Combining the features of being a boy, 14-years-old, with an indigenousfather,
and in a community with no secondary school means that a $200 conditional
transfer raises the probability of school enrollment by 23-24 percentage points.
These large differencessuggestthat there can be efficiencygains by usingsomeof
these dimensions of heterogeneity to target conditional transfers-in the same
way that Progresa uses gender differencesto set conditional transfer levels.
One concern is that identifying the impact of the size of a conditional transfer
on enrollment derives from observing children who are offered less than the full
transfer due to the cap on total household conditional transfers. These children
are, by definition, from households with more eligible children. To verify that
the enrollment model for these households does not differ significantlyfrom the
model for smaller households, the estimation was compared with a model
7. Another interesting result in column 3 is the relative magnitude of the impacts of a conditional
transfer (I,)compared with a nonconditional transfer (household total expenditure variable) on enroll-
ment. While the result is only suggestive because total expenditure is endogenous, [he $200 conditional
transfer is associated with an increase in the probability of enrollment 17 times higher than an equal
nonconditional transfer.
8. The age is centered on 12 years old, the median age for entry into secondary school, so that the
coefficient on the direct variable is the impact on a 12-year-old.
9. The lack of significance and very low point estimate of the maleStreatmentinteraction variable
(0.003 with standard error 0.002) are robust to many specifications, including both fewer and more
interaction terms and several control variables. The term is thus dropped from the estimation in column 5.
The often reported difference between boys and girls comes from estimations of enrollment rather than
continuation rates. Coady (2000)finds that most of that difference comes from Progress's very strong
impact on girls' re-entry into the school system during the program's fist year.
TAB LE 3. Heterogeneity: Impact of Conditional Transfers on the Probability of School Enrollment by Type of Child
Homogenous Impact I Ieterogenous impact
Treatment Conditional Transfer, Conditional Conditional Transfer
Community Dummy Conditional Transfer Amount with Control Transfer Amount with Control
Type of Child Variable (1) Amount (2) Variables (3) Amount (4) Variables (5)
Overall effect 0.130 0.140 0.140
Boy, 12 years old, 0.035 0.031
with nonindigenous
father, with secondary
school in the
!$ community ($200)
Boy, 14 years old
Boy, with
indigenous father
Boy, with no
secondary school in the
community
Boy, 14 years old,
indigenous father, with
no school in the
community
Column numbers correspond to those in Table 2.
Source: Marginal effects based on results from table 2.
de Janury and Sadoulet 17
estimated for these children alone. The estimation is, as expected, more precise
with the whole sample, but the parameters are neither individually nor globally
significantly different between the two estimations (the p-values for the test of
equality of the parameters on the conditional transfer variables are 0.49 without
heterogeneityand 0.16 with heterogeneity),which confirms that the conditional
transfer parameter was correctly identified.The orthogonality of the conditional
transfer to all other variables was also verified by estimatingdifferent models for
children in the treatment and control communities, and the parameters were
checked to ensure that they are neither individually nor globakly significantly
different between the two estimations. Thus, the model can be used to predict
behavior in the absence of a conditional cash transfer program.
IV. COMPARING CONDITIONAL CASH TRANSFER SCHEDULES
Three targeting and calibration schedules are now analyzed to determine
whether they can help raise the efficiency of conditional transfers in inducing
school enrollment (table 4). Each program has the same total budget as the
current Progresa program. This budget is calculated by predicting for each
sample child the expected uptake (predicted probability) EP and summing the
expected transfers EP.T to reach a total annual outlay of $322,000 for the 2,242
sample children.'' The upper panel of table 4 reports the enrollment rates for all
children and by category of children according to their 'risk level', that is, their
predicted enrollment rates without any conditional transfer, or their eligibility
status in the program. Thelower panel reports some aggregate targeting and cost
outcomes for the different schedules. These results are also represented in a
graph of the enrollment probability of each program against the initial enroll-
ment probability without a conditional transfer program (figure2).The distance
from the diagonal to each curve thus represents the gain in enrollment from the
program with the corresponding schedule.
Emulating Progresa: A Universal Uniform Conditional Cash Transfer Program
The school enrollment rate without conditional transfer is 63.2 percent [table 4,
column (I)].Progresa's current universal conditional transfers with a cap and
with differential values for boys and girls raise the probability of enrollment to
75.7 percent, a gain of 12.5 percentage points. The universal uniform condi-
tional transfers program without a cap and without gender differences used as a
benchmark for the subsequent simulations also raises the probability of enroll-
ment to 75.7 percent [column (2)].Under this program the conditional transfer
10. Another interesting exercise would be to define an efficient allocation of the total education
budget of the current Progresa program. It would consist of reallocating the primary school budget to
secondary school, thus doubling the budget for secondary school. A simulation of this budget reallocation
shows that it would lead to almost universal secondary education with enrollment rates of 90.491.7
percent, depending on the rule used for transfer calibration.
TABLE 4. Enrollment Rates under Alternative Conditional Cash Transfer Schedules
Predicted enrollment rates (percent)
Share of Universal Uniform Optimal Variable Implementable
Observations Total (percent) No Program Schedule Schedule Schedule
All children 2,242 100.0 63.2
Probability of enrollment
without conditional
transfer (percent)
0-40 354 15.8 27.8
40-60 583 26.0 50.9
60-70 376 16.8 64.9
70-80 392 17.5 74.6
80-100 537 24.0 90.5
r Eligible students
M
Without conditional transfer
With conditional transfer
Noneligible students
Eligibility (percent)
Average annual transfer amount
(US$)"
Annual cost per
additional child
enrolled (US$)
Efficiency gain over
universal uniform conditional
transfer schedule
(percent)
"Average over children who take the transfer.
Source: Authors' analysis based on the 2998 Progresa survey.
de Janvry and Sadoulet 19
FI RE 2. Impact of AlternativeConditional Cash Transfer Schedules on
GU
Enrollment Probabilities
Enrollment probability with conditional transfers
1.00 -
0.90 -
0.80 -
Optimal variable
0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 0.90 1.00
Enrollment probability without conditional transfer
Source:Authors' calculations based on the 1998 Progresa survey.
per child is about $194 a year.11 Because many children receive a transfer even
though they would attend school without one, the cost per additional child
enrolled is $1,151 a year. Figure 2 shows that gains in enrollment due to the
program are largest for children with a low probability of enrollment and decline
as the enrollment probability rises.
Table 4 reports these gains, with enrollment probability rising from an average
27.8 to 47.2 percent, or 19.4 percentage points, for the children with probability
of enrollment lower than 40 percent, while the gain is only 5.6 percentagepoints
for those in the 80-100 percent category [columns (1)and (2)].Gains are hence
progressive in terms of the initial likelihood of going to school, even with a
uniform conditional cash transfer program. This is the Progresa achevement
that has been widely acclaimed in the literature. However, can better results be
obtained by redefiningthe targeting and calibration of conditional transfers?
11. The levelofthe conditional transfer isdetermined by the Progresa budget, taking into account the
uptake that it is predicted to induce.
20 THE W O R L D BANK ECONOMIC REVIEW, VOL. 20, NO. I
An Optimal Variable Conditional Cash Transfer Program
The second simulation implements the optimal variable conditional transfer
program established in the model under the same budget constraint and taking
into account heterogeneity in probability of enrollment and responses to trans-
fers across children. Both eligibility and the optimal conditional transfer value
are simultaneously determined. This is done by offering the conditional transfer
defined in equation (7)to children of characteristics X. To compute the condi-
tional transfer values, the estimated values for /?, ij0,and 6in table 2, column (5),
are used, and the shadow value A of the budget constraint that balances the
budget is found by tgtonnement. The resulting conditional transfers vary from
$100 to $350, depending on the child's characteristics. Under this optimal
variable program the conditional transfers to children with a low probability
of attending school are increased, and children with high probabilities of attend-
ing school are targeted less because efficiency leakages are particularly high for
them.
The best predictor model for enrollment [table 2, column (5)]is again used to
predict enrollment for every child in the sample. Students eligible to receive a
conditional transfer have an average predicted enrollment rate of 78.9 percent,
compared with 55.8 percent, had they not been offered the conditional transfer.
The noneligible students have an average predicted enrollment rate of 89 per-
cent. Overall, the predicted school enrollment rate is now 81.1 percent, an
efficiency gain of 43.6 percent over the universal uniform conditional transfer
schedule. The optimal variable conditional transfer schedule almost equalizes
the probability of enrollment among children with very different initial prob-
abilities to values close to 70 percent (figure 2). The largest gains in probability
of enrollment are thus captured by those with the lowest initial probabilities.
Figure 3 shows the distribution of children by initial enrollment probability
without a conditional cash transfer program, superimposing the distribution of
those that are eligible in the optimal schedule (shaded)and showing, by differ-
ence, the distribution of noneligible (unshaded). Under the optimal conditional
transfer schedule eligibility is concentrated on children with low initial prob-
abilities, while noneligible children all have initial probabilities above 70 per-
cent. The optimal calibration of conditional transfers also favors those with low
initial probabilities, trying to induce them to enroll in school with a higher
conditional transfer. The conditional transfers decline as the probability of
being enrolled in school without a transfer rises. There are, however, relatively
few children with predicted low enrollment ~robability;most are in the 40-80
percent range.
Some 77.5 percent of children are eligible for a conditional transfer under the
optimal variable schedule, with the average transfer about $237, a 22 percent
increase over the average transfer under the universal uniform schedule (table 4).
The optimal schedule thus suggests raising the amount of the conditional transfer
for children who are less llkelyto attend school while reducing it for children who
are already likely to attend school without a conditional transfer. Since there are
de Janvry and Sadoulet 21
FIGURE 3 . Eligibility with the Optimal Variable Conditional Transfer
Schedule
Number of children
140 1
Enrollment probability without conditional transfer
Source:Authors' calculations based on the 1998 Progresa survey.
still efficiency leakages among eligible children, the cost per additional child
enrolled is $802, down from $1,151 under the universal uniform conditional
transfer schedule. Cost saving per additional child enrolled is around 30 percent.
An Implementable Conditional Cash Transfer Program
Having established the optimal conditional transfer schedule as an efficiency
benchmark, the analysis now turns to the definition of simpler implementable
conditional transfer schedules, based on a linear combination of a few observa-
ble characteristics. For a given set of variables Z the implementable schedule is
the solution to the optimization problem defined in equations (9)and (10).The
values for /3, aO,and 6 are taken from column (5)of table 2, with the parameters
cr: and /Z solved for iteratively,12 followed by exploration of combinations of
characteristics Z that are easily observable, verifiable by others, and nonmani-
pulatable by the household. An efficiencycriterion for selection requires choos-
ing characteristics that are important correlates of enrollment (to target the
children least likely to enroll without a conditional transfer) or that indicate
high sensitivity of enrollment to a conditional transfer. In addition to these
12. Starting with general eligibility, the optimization problem in equation 10 is solved for a as a
function of1,with1adjusted to balance the budget. These parameters are used to compute transfers and
define eligibility. This procedure is iterated until there is convergence, that is, no changc in eligibility
between two consecutive iterations. This is always achieved in fewer than five iterations.
22 THE WORLD BANK ECONOMIC REVIEW, VOL. 20, N O . I
features, actual implementation of a program requires these criteria to be legally
and politically acceptable. This is clearly an issue that every program should
address in its own particular context.
In the base model the conditional transfer schedule depends only on gender
and birth order of the child, presence of a secondary school in the community,
distance to a secondary school if there is not one in the community, and state
dummy variables, which are all strong correlates of enrollment. A few alterna-
tive specifications are reported later. Note that age of the child is not used
because an eligibility criterion based on age could induce perverse behavior,
such as parents delaying their children's entry to secondary school to benefit
from a larger conditional transfer. The child's birth order, which cannot be
manipulated, turns out to capture part of this information. Each variable can
be easily &served and verified. In fact, instead of secret eligibility formulas (as
currently used for poverty), which offer little room for recourse and account-
ability, self-registration is possible, with easy verification. .
Results are reported in table 5. The birth order parameter indicates that the
conditional transfer is highest for the oldest child and decreases by $12 for each
younger sibling. Girls would optimally receive a premium of $25. The main
source of variation in conditional transfer is related to distance to school, with a
large premium for children who need to travel some distance to school and an
additional amount for each kilometer traveled. The program also exhibits some
variation across states, with a difference of $87 between the extreme cases of
Queretaro and Guerrero.
Examples of eligibility and conditional transfer amounts computed with this
simple points system are reported in the lower part of table 5. Children with a
school in their own community (23percent of the sample) are not eligible. Their
predicted enrollment rate without conditional transfer is 80.5 percent, which is
also the rate observed in control communities with a school. By contrast, all the
children who do not have a school in their community are eligible for a condi-
tional transfer.13 A boy who is the oldest child and lives 3 kilometers from a
school would receive a conditional transfer of $213, whereas the third child
would receive only $190. If the oldest child is a girl, she would receive $239.
Cumulating all the disadvantages, a girl who is the oldest child and lives 6
kilometers from a school would receive the highest transfer: $266.
This conditional transfer schedule results in an efficiency loss relative to the
optimal variable schedule, the cost to be paid for simplicity, and transparency
(table4).Although the number of eligible children is about the same as with the
optimal conditional transfer schedule (77.4 percent), the eligibility criteria are
not the same. The implementable schedule includes 9 percent of the children not
13. The average distance to school for the 77 percent of children who do not have a school in their
community is 3.1 kilometers. Enrollment rates decrease very sharply with distancc to school in the control
communities, reaching a low of 43 percent for the 19 percent of children who live more than 4 kilometers
from school.
de Janvry and Sadoulet 23
TABLE 5. Optimal Implementable Conditional Cash Transfer Schedule
Base Model With Illiteracy Geographic
Conditional transfer formula
(US$per year)
Birth order -12 -12
Male -25 -25
No secondary school 476 502 447
in the community
Distance to secondary 50 49 48
school [In(l+kms)]
Mother illiterate 26
Father illiterate 30
Guerrero -295 -351 -295
Hidalgo -278 -327 -283
Michoacan -246 -288 -253
Puebla -267 -317 -273
Queretaro -208 -257 -214
San Luis Potosi -260 -303 -260
Veracruz -285 -333 -288
Examples of transfers
by children types in state
of Guerrero (US$ per year)
School in the community Not eligible Not eligible Not eligible
Oldest, male, with literate 213 184 218
parents, and school 3 kilometers
away
Third child, male, with literate 190 162 218
parents, and school 3 kilometers
away
Oldest, male, with illiterate parents, 213 240 218
and school 3 kilometers away
Oldest, female, with illiterate parents, 239 265 218
and school 3 kilometers away
Oldest, female, with illiterate parents, 266 292 245
and school 6 kilometers away
Efficiency gain over universal uniform 29.4 31.0 28.5
schedule (percent)
eligible under the optimal schedule (type I1error) and excludes 9 percent of the
children eligible under the optimal schedule (typeI error). Enrollment of eligible
children rises from 58.2 percent without a conditional transfer to 79.1 percent
with one. Enrollment is 80.5 percent for noneligible children and 79.4 percent
for all children. This implies a 29.4 percent efficiency gain over the universal
uniform conditional transfer schedule. Cost per additional child enrolled is
$889, still about 23 percent cheaper than under the universal uniform schedule
but nearly 11 percent more expensive than under the optimal variable schedule.
Alternative implementable schedules that vary the characteristics used
to establish eligibility and conditional transfer amounts are now explored
(table 5). Adding mother's and father's literacy status, important predictors of
school enrollment, raises the efficiency gain to 31 percent above the universal
uniform schedule. While it can be argued that such subsidies (herecomputed as
$26 if the mother is illiterate and $30 if the father is illiterate) may give the
. -
wrong signal and bias, the return to education, they can also be seen as a way to
compensate for the handicap that children of uneducated parents have and to
help children catch up.
At the other extreme, how efficientwould it be to define a conditional transfer
schedule at the communitylevel (althoughonlyfor poor children)?Doing so leads
to an important efficiency gain of 28.5 percent over the uniform conditional
transfer schedule. This geographical targeting is interesting because it shows that
in the particular case of rural Mexico an important efficiency gain could be
obtained by redesigningthe conditional transfer program as a school transporta-
tion subsidy. A simple transportation subsidy would capture 65 percent of the
efficiency gain that the optimal conditional transfer program would garner. The
question then arises of how this intervention compares with a supply-side policy
that would bring schools closer to where people live. This is beyond the scope of
this article, but Coady (2000)estimates that the cost of raising enrollmentthrough
a supply-side intervention that increases the number of rural schools would be
more than seven times as much as the current Progresa program.
These specific implementable schedules are illustrations of the idea that
designing a relatively simple conditional transfer program, with a points system
that is transparent and easily verifiable, is indeed feasible and could ensure large
efficiency gains.
Comparing Direct Costs and Efficiency Leakages under the Three Schedules
A key determinant of the relative efficiency of different targeting schedules is
their efficiency leakages, namely the magnitude of the transfers that go to
children who would attend school without the conditional transfer. This is
analyzed in a comparative fashion in figures 4 and 5, where the total transfer
cost for each category of children is divided into direct costs (transfers to
children that would not otherwise have enrolled, represented in black) and
efficiency-leakage costs (transfersto children that would have enrolled anyway,
represented with stripes). Differences among the figures are quite telling. Under
the universal uniform conditional transfer schedule, leakages are particularly
high, especially among children with a high probability of attending school
without a conditional transfer. Some 83.2 percent of the total cost is efficiency
leakages, leaving an effective direct cost of only 16.8 percent. The optimal
variable conditional transfer schedule reduces efficiency leakages by targeting
children with a low probability of attending school and increasingthe magnitude
of the conditional cash transfers (figure 5). Efficiency leakages are reduced to
64.9 percent of total costs, implying an effective direct cost of 36.1 percent. The
implementable conditional transfer schedule (not pictured) has an efficiency
leakage of 72.5 percent of total costs. Because targeting is simplified and
de lanury and Sadoulet 25
FIGURE 4. Total Annual Direct Costs and Leakage Costs under the Universal
Uniform Conditional Transfer Program
US$ (thousands)
400 1
300 - Efficiency leakage costs
Overall share = 83.2%
250 -
ElEfficiency leakagecosts
0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
Enrollmentprobability without conditional transfer
Source: Authors' calculations based on the 1998 Progresa survey.
transparent, it is a compromise between the universal and the optimal variable
conditional transfer schedules. Its effective direct cost is 27.5 percent.
In conclusion, the optimal variable conditional transfer schedulecould offer a
significant efficiency gain in school enrollment and could be implemented
through a secret formula like the one Progresa currently uses to target poverty.
It may, however, be too complex to administer, and secrecy is not a desirable
feature as it makes recourse almost impossible. However, the implementable
variable conditional transfer schedule-with its transparent targeting-also
results in substantial efficiency gains relative to Progresa's current universal
uniform conditional transfers.
Are these optimal and implementableschedules regressiveor progressive?In other
words, are efficiency gains in enrollment achieved at an equity cost? Conditional
transfers driven by efficiency gains indeed raise the issue that maximally efficient
programs may be inequitable (Das,Do, and Ozler 2005).For this reason eligibility
is restricted to poor households. However, when poor households are further
targeted, are the resultingtransfers regressive or progressive?
Before looking at the distributive effect of targeting among poor households,
it is interesting to note that the Progresa transfers themselves were not
FIGURE 5 . Total Annual Direct Costs and LeakageCosts under the Optimal
Variable Conditional Transfer Program
US$ (thousands)
1
iEfficiency leakage costs
Overall share= 64.9%
0.05 0.15 0.25 0.35 0.45 0.55 0.65 0.75 0.85 0.95
Enrollmentprobability withoutconditionaltransfer
Source: Authors' calculations based on the 1998 Progresa survey.
particularly efficient in reducing poverty or inequality. Indeed, with poverty
measured by consumption per capita, the transfers are almost uniformly
distributed across levels of per capita consumption (de Janvry and Sadoulet
2003). This article, however, discusses the issue of tradeoff between efficiency
and equity using the Progresa welfare index measured in 1997, rather than the
income-consumption level, since this is what Progresa uses as a poverty indica-
tor. Figure 6 shows the average distributed transfer with households ranked by
the Progresa welfare index. The average transfer distributed by Progresa shows a
clear upward trend and thus regressivityamong poor households.This is because
of the low uptake rate in low welfare classes. By contrast, the average distributed
transfer decreases across welfare levels with the optimal variable conditional
transfer schedule (from $160 to $140) and is uniform with the implementable
schedule. Efficiency gains in implementing conditional transfer programs
designed to maximize the effect of the conditionality are thus not achieved at
the cost of rising inequality among poor households.
VI. CONCLUSIONS
This article questions whether efficiency gains can be achieved in conditional
cash transfer programs by improving targeting among poor households and
de Janvry and Sadoulet 27
FIGURE 6. Eligibility and Distributed Transfers by Welfare Index
Average annual distributedtransfer (US$)
Optimal variable
120 1 I I I I I I I r
0 10 20 30 40 50 60 70 80 90 100
Percentile of welfareindex
Source: Authors' calculations based on the 1998 Progresa survey.
better calibrating conditional transfers. The efficiency objective is to maximize
the impact of the condition imposed on the transfer, in this case gains in school
enrollment among poor children. Using the data from the Progresa randomized
experiment, the analysis focused on the crucial education decisionfor children in
poor Mexican rural communities-namelywhether to continue schooling at the
secondary level.
Achieving efficiency gains by targeting and calibrating condtional transfers
requires focusing on children who have a high probability of not enrolling in
school without a conditional transfer and who have a high response to the amount
offered, within the overall program budget constraint. Implementing this program
requires predicting school enrollment as a function of the conditional transfer
offered and of child, household, and community characteristics.Heterogeneityin
responses shows that age, ethnicity, and presence of a school in the community
lead to large differencesin enrollment. Three alternative targeting and calibration
scheduleswere then compared: the current ~ r o ~ r eschedule of universal uniform
s a
conditional transfers; an optimal schedule of variable conditional transfers; and a
schedule of implementable conditional transfers, where the criteria used for
targeting and calibration are easily observable, verifiable by others, and nonrna-
nipulatable by the household. In setting up a new program, a pilot experiment
would need to be used to estimate the enrollment probability model necessary to
establish the targeting and calibration formulas.
The optimal schedule is found to offer a 44 percent efficiency gain over the
universal conditional transfer schedule and the implementable schedule a 29
percent gain. The optimal schedule reduces efficiency leakages (receipt of trans-
fers by children who would attend school without a conditional transfer) from
83 percent of total costs to 65 percent, and the implementable schedule to 73
percent. These efficiency gains are not achieved at the cost of rising inequality
among poor households.
The overall conclusion is thus that large efficiency gains can be achieved in
implementing what are in many countries very expensive conditional cash
transfer programs for human capital formation among poor households if
rules for targeting and calibration of conditional transfers are designed to max-
imize the effect of the conditionality.
Alderman, Harold. 2001. "Multi-Tier Targeting of Social Assistance: The Role of Intergovernmental
Transfers." World Bank Economic Review 15 (1):33-53.
2002. "Do Local Officials Know Something We. Don't? Decentralization of Targeted Transfers in
Albania." Journal of Public Economics 83 (3):307457.
Behrman, Jere, and Petra Todd. 1999. "Randomness in the Experimental Samples of PROGRESA." Interna-
tional Food Policy Research Institute, Washington, D.C.
Behrman, Jere, Piyali Sengupta, and Petra Todd. 2001. "Progressing Through Progresa: An Impact
Assessment of a School Subsidy Experiment." University of Pennsylvania, Department of Economics
Philadelphia.
Bourguignon, Frangois, Francisco Ferreira, and Philippe Leite. 2002. "Ex.-Ante Evaluation ofCCTPrograms:
The Case of Bolsa Escola." Policy Research Working Paper 2916. World Bank, Washington, D.C.
Coady, David. 2000. "The Application of Social Cost-Benefit Analysis to the Evaluation of Progresa."
International Food Policy Research Institute, Washington, D.C.
Das, Jishnu, Quy-Toan Do, and Berk Ozler. 2005. "Reassessing Conditional Cash Transfer Programs."
World Bank Research Observer 20 (1):57-80.
Gertler, Paul. 2004. "Do Conditional Cash Transfers Improve Child Health? Evidence from Progress's
Control Randomized Experiment." American Economic Review 94 (2):33641.
deJanvry, Alain, and Elisabeth Sadoulet. 2003. "Dkpasser Bono: Comment rendre plus efficiente I'aide au
diveloppement." Revue D'iconomie Du Diveloppement 4:63-76.
de Janvry, Alain, Frederico Finan, and Elisabeth Sadoulet. 2001. "Decomposing the Channels of Influence
of Ccts in a Structural Model of Educational Choice." University of California at Berkeley. http://
are.Berkeley.edu/-sadouletl
Morley, Samuel, and David Coady. 2003. "From Social Assistance to Social Development: A Review of
Targeted Education Subsidies in Developing Countries." Center for Global Development and Inter-
national Food Policy Research Institute, Washington, D.C.
Ravallion, Martin. 2003. "Targeted Transfers in Poor Countries: Revisiting the Trade-offs and Policy
Options." Policy Research Working Paper 3048. World Bank, Washington, D.C.
de Janvry and Sadoulet 29
Ravallion, Martin, and Quentin Wodon. 2000. "Does Child Labor Displace Schooling? Evidence on
Behavioral Responses to an Enrollment Subsidy." Economic Journal 110 (462):158-75.
Rawlings, Laura, and Gloria Rubio. 2005. "Evaluating the Impact of Conditional Cash Transfer Pro-
grams." World Bank Research Observer 20 (1):29-55.
Schady, Norbert. 2002. "Picking the Poor: Indicators for Geographic Targeting in Peru." Review of
Income and Wealth 48 (3):417-33.
Schultz, T. Paul. 2004. "School Subsidies for the Poor: Evaluating the Mexican Progresa Poverty Pro-
gram." Journal of Development Economics 74 (1):199-250.
Skoufias, Emmanucl. 2000. "Is PROCRESA Working? Summary of the Results of an Evaluation by IFPRI."
FCND Briefs International Food Policy Research Institute, Food Consumption and Nutrition Division,
Washington, D.C.
van de Walle, Dominique. 1998. "Targeting Revisited." World Bank Research observe^ 13 (2):23148.
Child Labor and School Achievement
in Latin America
Victoria Gunnarsson, Peter F. Orazem, and Mario A. Shnchez
Child labor's effect on academic achievement is estimated using unique data on third
and fourth graders in nine Latin-American countries. Cross-country variation in tru-
ancy regulations provides an exogenous shift in the ages of children normally in these
grades, providing exogenous variation in the opportunity cost of children's time. Least
squares estimates suggest that child labor lowers test scores, but those estimates are
biased toward zero. Corrected estimates are still negative and statistically significant.
Children working 1standard deviation above the mean have average scores that are 16
percent lower on mathematics examinations and 11 percent lower on language exam-
inations, consistent with the estimates of the adverse impact of child labor on returns
to schooling.
About one of eight children in the world is engaged in market work. Despite
general acceptance that child labor is harmful and despite international accords
aimed at its eradication, progress on lowering the incidence of child labor has
been slow. Although often associated with poverty, child labor has persisted in
some countries that have experienced substantial improvements in living stan-
dards. For example, Latin America, with several countries in the middle- or
upper-middle-income categories, still has child labor participation rates that are
similar to the world average.
Countries have adopted various policies to combat child labor. Most have
opted for legal prohibitions, but these are only as effective as the enforce-
ment. As many child labor relationships are in informal settings within
family enterprises, enforcement is often difficult. Several countries, particu-
larly in Latin America, have initiated programs that offer households an
income transfer in exchange for keeping children in school and out of the
labor market.
Victoria Gunnarsson is a research officer in the Fiscal Affairs Department at the International
Monetary Fund; her email address is vgunnarsson@irnf.org. Peter F. Orazem is a professor at Iowa
State University; his email address is pfo@iastate.edu. Mario A. SLnchez is a social development specialist
at the Inter-American Development Bank; his e-mail address is mariosan@iadb.org. The authors thank
Wallace Huffman, Robert Mazur, three referees, the editor, and seminar participants at Iowa State and
Minnesota for numerous comments and suggestions.
THE WORLD BANK ECONOMIC REVIEW, VOL. 20, NO. 1, pp. 31-54 doi:10.1093/wber/lhj003
Advance Access publication March 2, 2006
0TheAuthor 2006. Published by OxfordUniversityPresson behalf of the International
Bank for Reconstructionand Development/ THEWORLD BA NK.All rights reserved. For permissions,
please e-mail: journals.permissions@oxfordjournals.org.
Presumably, governments invest resources to lower child time in the labor
market in anticipation that the child will devote more time to the acquisi-
tion of human capital. The government's return will come from higher
average earnings and reduced outlays for poverty alleviation when the
child matures. However, despite a huge acceleration in the research on
child labor, there is surprisingly little evidence that relates child labor to
schooling outcomes in developing countries.' Most children who work are
also in school, suggesting that child labor does not lower school attainment.
Additionally, studies that examine the impact of child labor on test scores
have often found negligible effects, although most of these are in industrial
country contexts. More recently, Heady (2003)and Rosati and Rossi (2003)
have found some evidence that child labor lowers primary school test scores in
developing countries.
This article builds on these last two studies by examining- the link between
child labor and school achievement in nine countries in Latin America. This
article benefitsfrom more detailed data sets that allow controls for child, house-
hold, school, and community variables, and it uses an empirical strategy that
controls for the likely endogeneityof child labor. The results are consistent: in all
nine countries, child labor lowers performance on the tests of language and
mathematics proficiency, even when controlling for school and household attri-
butes and for the joint causality between child labor and school outcomes. To
the extent that lower cognitive attainment translates into lower future earnings,
as argued by Glewwe (2002),these results suggest that there is a payoff in the
form of higher future earnings from investing in lowering the incidence of child
labor.
Most studies that analyze the relationship between time at work and school
attainment have focused on high-school or college students in industrial coun-
tries.2 These studies have generally found little evidence that part-time work
combined with schooling affects school achievement. When adverse effects are
found, they are apparent only at relatively high work hours. Important excep-
tions include recent studies by Tyler (2003)and Stinebrickner and Stinebrick-
ner (2003)that found that after controlling for the likely endogeneity of child
labor, working while in school led to much larger implied declines in high-
school math scores and in college grade point averages than had been found
previously. Post and Pong (2000) also found a negative association between
1. Two excellent recent reviews of the recent literature are by Basu and Tzannatos (2003) and
Edmonds and Pavcnik (2005).
2. D'Amico (1984),Ehrenberg and Sherman (1987),Howard (1998),Lillydahl(1990),Singh (1998),
Stern (1997),and Singh and Ozturk (2000).
Gunnarsson, Orazem, and Sanchez 33
work and test scores in samples of eight graders in many of the 23 countries
they ~tudied.~
There are several reasons why the experience of older working students may
differ from that of young children working in developing countries. Young
children may be less physically able to combine work with school, so that
young working children may be too tired to learn efficiently in school or to
study afterwards. Children who are tired are also more prone to illness or injury
that can retard academic development. It is possible that working at a young
age disrupts the attainment of basic skills more than it disrupts the acquisition
of applied skills for older students. School and work, which may be comple-
mentary activities once a student has mastered literacy and numeracy, may not
be compatible before those basic skills are mastered.
Past research on the consequences of child labor on schooling in developing
countries has concentrated on the impact of child labor on school enrollment or
attendance. Here the evidence is mixed. Patrinos and Psacharopoulos (1997)and
Ravallion and Wodon (2000)found that child labor and school enrollment were
not mutually exclusive and could even be complementary activities. However,
Rosenzweig and Evenson (1977) and Levy (1985) found evidence that better-
developed child labor markets lowered school enrollment. There is stronger evi-
dence that chld labor lowers time spent in human capital production, even if it
does not lower enrollment. Psacharopoulos (1997) and Sedlacek and others
(forthcoming)reported that child labor lowered years of school completed, and
Akabayashi and Psacharopoulos (1999) discovered that child labor lowered
study time.
Nevertheless, school enrollment and attendance are not ideal measures of the
potential harm of child labor on learning because they are merely indicators of
the time input into schooling and not the learning outcomes. Even if child labor
lowers time in school, it may not hinder human capital production if children
can use their limited time in school efficiently. This is particularly so if schools
are of such poor quality that not much learning occurs in any case. By contrast,
the common finding that most working children are enrolled in school may miss
the adverse consequences of child labor on learning if child labor is not com-
plementary to the learning process at the lower grades.
A more accurate assessment of the impact of child labor on human capital
production requires the measures of learning outcomes, such as test scores,
rather than education inputs, such as time in school, to determine whether
child labor limits or enhances human capital production. Moreover, evidence
suggests that cognitiveskills, rather than years of schooling, are the fundamental
determinants of adult wages in developingcountries (Glewwe1996, Moll 1998).
3. The study included several developingcountries, including Colombia, Iran, South Africa, Thailand,
and the Philippines, which had the largest estimated negative effectsof child labor on school achievement.
However, the estimates do not control for school attributes or possible joint causality between school
achievement and child labor.
3 4 THE W O R L D BANK E C O N O M I C REVIEW, VOL. 20, N O . I
Therefore, identifyingthe impact of child labor on school achievementwill yield
more direct implications for child labor's longer-term impacts on earnings and
poverty status later in the child's life.
Direct evidence of the impact of child labor on primary school achievementis
rare. Heady (2003)found that child work had little effect on school attendance
but a substantial effect on learning achievement in reading and mathematics in
Ghana. Rosati and Rossi (2003)reported that in Nicaragua and Pakistan, more
hours of child labor are associated with poorer test scores. Both of these studies
have weaknesses related to data limitations. Heady treated child labor as exo-
genous, but it is plausiblethat parents send their children to work in part because
of poor academic performance. Rosati and Rossi had no information on teacher
or school characteristics, although these are likely to be correlated with the
strength of local child labor markets.
This study makes several important contributions to knowledge of the
impact of child labor on schooling outcomes in developing countries. It
shows how child labor affects test scores in nine developing countries, greatly
expanding the scope of existing research. Because the same examination was
given in all countries, the study can illustrate how the effect of child labor on
cognitive achievement varies across countries that differ greatly in child labor
incidence, per capita income, and school quality. Because the countries also
differ in the regulation and enforcement of child labor laws, cross-country
variation in schooling ages and truancy laws can provide plausible instruments
for endogenous child labor. Finally, because the data set includes a wealth of
information on parent, family, community, and school attributes, the impact of
child labor on schooling outcomes can be estimated while holding fixed other
inputs commonly assumed to explain variation in schooling outcomes across
children.
The results are consistent. Child labor lowers student achievement in every
country. The conclusions are robust to alternative estimation procedures and
specifications. The inescapable conclusion is that child labor has a significant
opportunity cost in the form of forgone human capital production, a cost that
may not be apparent when looking only at enrollment rates for working
children.
Ben Porath (1967)laid out the classic model of human capital investments over
the lifecycle. There are diminishingmarginal returns to time in school because of
concavity in the human capital production process and because the opportunity
cost of allocating time to further skill acquisition increases as skills are accumu-
lated. In addition, finite life spans limit the length of time to capture returns from
schooling as age increases, further decreasing the marginal returns to time in
school as age rises. All of these factors suggest that time invested in human
capital production will decrease as an individual ages. However, early in life,
Gunnarsson, Orazem, and Sa'nchez 35
children may specialize in schooling if the present value of the return is suffi-
ciently high relative to its current marginal cost.4
Of interest here is the tradeoff parents face in deciding whether a child should
specialize in schooling or should divide time between school and work. By age t,
the child has completed Etyears of schooling. In addition, the child has matured
for t years. The opportunity cost of a child's school time is assumed to rise with
Etand t and is also a function of local labor market conditions Z,. The returns to
time in school will depend on how much the child is expected to learn, Q,. A
vector of observable parent, home, school, and community variables, H,, may
affect tastes for child labor as well as the productivity of child time in school
through Q,. The child's labor supply function will be of the form
Ct = 4% t,Zt,Qt, Ht, ~ t )
where E~is a random error.
The human capital production process is assumed to depend on past human
capital accumulations, current factors that would make the child's time in school
more productive, and the time spent in school. Letting Q, be an observable
measure of cognitive skills produced in school, the human capital production
process will be of the form
where q, is a component of cognitive ability that the parents can observe but not
the econometrician.
Because the decision on whether or how much the child works is based in part
on the parents' knowledge of q , and because student outcomes are influenced by
child labor, Var(~,,qt)# 0, and ordinary least squares estimation of equation (2)
will be biased. Short of a randomized experiment that assigns children into
working and non-working groups, the best candidate to resolve the problem
will be to find variables that shift the probability that a child works but do not
directly affect child learning in school. Needed are variables that alter the local
labor market for child labor, Z,, to provide exogenous shifts in the child labor
equation in estimating equation (2).
Factors Shifting the Probability of Child Labor
Elements of the vector Z, are required that alter the local labor market for
children but do not affect test scores. Because the probability of working rises
4. Themain predictions are not altered if leisureisadded to the model. It willstill beoptimal to invest
more intensively in human capital early in life and to decrease investment intensity with age. In addition,
because the cost of leisure is the value of work time, individuals will consume the least leisure when wages
are highest. In the application here, children will consume less leisure as they age, and so older children
will still be expected to work more than younger children. Heckrnan (1976)presented a detailed model of
human capital investment, leisure demand, and consumption over the life cycle. Huffman and Orazem
(2006)present a much-simplified model that generates the predictions discussed in the text.
with age, factors that alter the age at which a child would normally be in a given
grade will also affect the probability that the child will be working. In Latin
America, the age at which children are expected to start school varies across
countries from 5 to 7 years of age. The age at which a child may legally leave
school also varies from 12 to 16 years of age. As a consequence, children must
attend school as few as 5 years in Honduras to as many as 10 years in Peru.
These differences in laws regulatingschool attendance and child labor alter the
age at which children would normally enter grades 3 and 4 and thus the opportu-
nity costs of beingin thosegrades. Children starting school earlier will be younger
at grade 3 and more likely to attend school full time without working. Third and
fourth gradersin countries with the lowest working ages are more likely to appear
legal, even if they are under 12 years of age. Therefore, children in countries with
low truancy ages will be more likely to be working while attending school.
An alternative measure of the opportunity cost of attending school would be
the local market wage for children. Because most child labor is unpaid work for
family enterprises, however, market wages would not adequately capture the
value of time outside of school even if such information were available. In their
place is used the presumed upward relationship between the marginal produc-
tivity of child labor and the child's age, assumed to be driven largely by physical
stature.' Interactions between measures of a country's school starting age or
truancy age and a child's age are used to capture exogenous variation across
countries in the probability that third and fourth graders work. These shifts in
the net return to time in school provide the needed exogenous shift in c . ~
Within countries the largest source of variation in demand for child labor
occurs across rural and urban areas. There are more uses for child labor in rural
markets, and so labor force participation rates are higher for rural children than
for urban children in all the countries in this study. That source of variation is
captured with interactions between child age and a dummy variable indicating
rural residence for boys and girls.
How these elements of Z, affect the probability of engaging in child labor is
illustrated in figures 1-3.
Factors Affecting School Outcomes
Estimation of equation (2)follows the educational production function litera-
ture in that Q is measured by test scores that are explained by variables char-
acterizing the student's parents, household, teacher, school, and community
5. Rosenzweig (1980)found that in a sample of adults, wages for day labor in India were primarily
driven by stature and not by acquired education. Wage patterns reported by Ray (2000)for boysand girls
in Pakistan and Peru suggest rising opportunity costs of child time as age increases.
6. Angrist and Krueger (1991) used variation in compulsory school starting ages across states to
instrument for endogenous time in school in their analysis of returns to schooling using U.S. Census data.
Tyler (2003)used variation in state child labor laws to instrument for child labor in his study of U.S. high-
school test scores. This study began with a large number of interactions, but the resulting variables were
highly collinear, and so a parsimonious subset of the fuller specification was used.
Gunnarsson,Orazem, and Sknchez 37
FI GURE 1. Predicted Child Labor by Child Age and School Starting Age
Source:Authors' simulations based on results in table 2, column 1.
(Hanushek 1995). Measures used include most of those that have been found to
be important in developing country settings (Hanushek 1995, Kremer 1995).
Estimates of educational production functions are subject to numerous
bia~es.~Among the most commonly discussed is the lack of adequate control
for the student's innate ability.8Many studies have attempted to correct for the
problem by using two test scores taken at different times. If ability has an
additive effect on school achievement, the difference between the two output
measures will be purged of the ability effect. The data for the current study
include only tests taken at one point in time, so the differencing option is not
available. However, there are reasons why undifferenced data may yield satis-
factory or even preferred estimates to the differenced data. As Glewwe (2002)
argued, if measures of H, vary slowly over time, the value of the differenced
measure of achievement is minimal. This is more likely to be true at the earliest
stages of schooling, when there is less variation in curriculum, educational
materials, or teacher training. Furthermore, the use of parental attributes such
7. See Glewwe (2002)for a comprehensive reviewof the problems associated with estimating educa-
tional production functions.
8. Ability bias has also been the subject of numerous studies estimating returns to schooling. The
consensus is that the bias is small (Card1999). If earnings and cognitiveskillsare closely tied, as argued by
Glewwe (2002),the role of ability bias should be small in educational production estimates also.
FIGURE 2. Predicted Probability of Child Labor by Child Age and School
Leaving Age
--
Source: Authors' simulations based on results in table 2, column 1.
as education and income should partially control for inherited ability. Finally, if
there is considerable measurement error in estimates of Q,, the level of Q, may
be measured more reliably than the change in Q,. In any event, the results of the
production function estimation in this study should be interpreted as cumulative
as of grade 3 or 4 rather than the additional learning obtained in that grade.
In 1997 the Latin-American Laboratory of Quality of Education (LLECE)carried
out the First Comparative International Study on Language, Mathematics, and
Associated Factors for third and fourth graders in Latin America. LLECE initially
collected data in 13 countries, but the required information for the regression
analysis for this study was available only for nine countries: Argentina, Bolivia,
Brazil, Chile, Colombia, Dominican Republic, Honduras, Paraguay, and ~ e r u . ~
The data set is composed of a stratified sample designed to ensure sufficient
observations of public, private, rural, urban, and metropolitan students in each
9. Costa Rica was included in the initial data collection, but LLECE dropped those data because of
consistency problems. Cuba was excluded because of missing data on child labor. Mexico and Venezuela
lacked required information on child age.
Gunnarsson, Orazem, and Sanchez 39
FIGURE 3 . Predicted Child Labor Probability by Child Age, Gender, and
Region
' - 1
Ruralboys 1
4 Metropolitanboys
Q 10 11 12 13 14 15 i 6
Age
Source: Authors' simulations based on results in table 2, column 1.
country. Data were collected on 40 chldren from each of 100 schools in each
country for a total of 4,000 observationsper country. Half of the students were in
the third grade and half in the fourth grade. For budgetary reasons LLECE had to
use a priori geographicexclusionsto limit the transportation and time costs of data
collection. Very smallschools with too few thrd and fourth graders and schools in
remote, difficult to access, or sparselyinhabited regionswere excluded. Because of
the cost of translatingexaminations,schoolswith bilingualor indigenouslanguage
instructionwere also excluded.1� As the excluded schoolswould cater to relatively
more disadvantaged populations, our results should be viewed as applying to
school populations that are less rural, from more majority ethnic groups, and
somewhat more advantaged than average for all Latin-American chldren.
Test Scores
Survey instrumentsconsisted of tests administeredto the sample of children of the
sampled schools, and self-applied questionnaires to school principals, teachers,
parents (or legal guardians) of the tested children, and the children. In addition,
surveyors collected information on the socioeconomic characteristics of the
10. For a detailed description of the a priori exclusions in each country, see Table 111.6 of the
Technical Bulletin of the LLECE.
community. A description of the variables used in the analysis is provided in
appendix table A-1, and summary statistics are reported in appendix table A-2."
All children were tested in mathematics, and all were tested in Spanish except
the Brazilian children who were tested in Portuguese. The tests and question-
naires were given only to children who attend school, so no information was
obtained on children who are not in school. Therefore, the results can be applied
to enrolled children only. If working children who perform most poorly in
school drop out to work full time, the estimate of the consequences of child
labor on schooling outcomes may miss some of those most harmed by child
labor while including children who can work and still perform well in school.
However, 95 percent of children aged 9-11 are enrolled in Latin America, so the
bias is likely to be modest.12In settingswhere primary enrollment rates are much
lower, the bias could be substantial, however.
Child Labor
Child labor is measured by children's responses to a question asking whether they
are engaged in work outside the home.13The concentration on paid work outside
the home avoidssome definitionalproblems related to distinguishingunpaid work
for home enterprise from household chores. However, it is also apparent in the
application that child labor in the home does not have the same apparent negative
consequenceson student achievement as does work outside the home.
A comparisonof the intensity of child labor participation rates in nine countries
for children who report that they work inside or outside the home and average
language and mathematics test scores shows an unvarying pattern (table l).14
Children who work only some of the time outperform those who work often.
Children who almost never work outperform those who work sometimesor often.
The differences are almost always statistically significant. The advantage is large
for children who almost never work over those who often work, averaging
22 percent on the mathematics examination and 27 percent on the language
11. For some reason, language scores were reported for 2 percent fewer students than were mathe-
matics scores. The missing scores appear to be due to random reporting errors, as there were no large
differences between the sample means of the group taking the mathematics and language tests. The means
are reported from the sample taking the mathematics examination.
12. Sedlacek and others (2005)presented data on enrollment by age for 18 Latin-American countries.
Even for the poorest quintile of children, enrollment rates are more than 90 percent for children aged 9-11.
13. As pointed out by a referee, it would be better to have information on hours of work rather than
these more-vague measures of work intensity. The instrumental variables procedure described later is an
attempt to correct for biases because of measurement error in child labor.
14. The averages are reported for the subset of countries for which data were available on both
language and mathematics test scores and for which responses could be matched for working inside and
outside the home. Only partial information was available for Mexico and Venezuela, but the pattern of
average test scores for children working outside the home in Mexico and Venezuela was the same-
children working more outside the home had significantly lower average test scores. Data limitations
prevented generating the corresponding average test scores for children working inside the home for those
two countries.
A
F4
m
Y , k l
u
3 .
u
4
U
2
TY A
m
*
II
9 , s e
FS c
g,S"
E
TAB LE 1. Continued
Working Outside the Home Working Inside the Home
Language Mathematics Language Mathematics
Test Scores Test Scores Test Scores Test Scores
Country (Maximum Score = 19) (Maximum Score = 32) (Maximum Score = 19) (Maximum Score = 32)
Honduras
Often
Sometimes
Almost never
Paraguay
Often
Sometimes
Almost never
Peru
Often
Sometimes
Almost never
All countries
Often
Sometimes
Almost never
"Differencefrom often working group significant at the 0.05 confidence level.
""Diffexence from often working group significant at the 0.01 confidence level.
Note: Results are the simple mean test score over all children in the child labor group in the county. Numbers in parentheses are the percentage difference
relative to children who often work outside the home when not in school. For definitions of "often," "sometimes," and "almost never," see tab1e.A-1.
Source: Authors' computations based on data from the 1997 survey by the Latin-American Laboratory of Quality of Education, as described in the text;
UNESCO (2002).
Gunnarsson, Orazem, and Sdnchez 43
examination. The test advantage for occasional child laborers is smaller but still
significant at 8.4 percent for mathematics and 9 percent for languages.
Children were asked a similar question about how intensively they worked
inside the home. It seems that working inside the home is less costly for human
capital development in schools. Across all countries, those who work often
inside the home have average test scores only 7 percent lower than those who
almost never work inside the home and only 4 percent lower than those who
sometimes work inside the home. The test score gaps for those working outside
the home were considerably larger. Furthermore, in only three of the nine
countries were average test scores significantly higher for children almost
never working inside the home relative to those often working inside the
home. In three other countries, those often working inside the home had higher
average test scores than did those rarely working inside the home.
Nevertheless, there is a more basic reason for not analyzing the implications
of working inside the home on student achievement: more than 95 percent of
students reported working inside the home sometimes or often, with nearly
identical incidence of work reported for girls and boys and for urban and rural
children. This lack of meaningful variation means that the pattern of test scores
against work intensity inside the home is unlikely to be reliable. In fact,
attempted empirical models could not distinguish statistically between children
who did and those who did not work inside the home-everyone was predicted
to participate in household labor. It is possible that Gork inside the home is
damaging to schooling outcomes, but our data lack sufficient variation in
measured household work to capture the effect. For these reasons, we concen-
trate our analysis on child labor outside the home.
Exogenous Variables
The presumed positive relationship between age and the value of child time
working outside the home is used to identify the child labor equation. This
relationship varies across urban and rural areas and between boys and girls. It
also appears to shift as children reach 10 years of age. This effect is allowed with
a spline defined as follows. A dummy variable, dlO, takes the value of 1 for
children under 10 years of age and 0 otherwise. For children aged 10 and older,
the age effect is captured by interactions between (1- d10) and age.
The countries included in the data differ in their legal regulations governing
the age at which children enter school and when they can leave school. Informa-
tion on compulsory schooling laws for each country was obtained from the
UNESCO (2002).In the empirical specification, these laws shift the age-child
labor relationship beyond age 10, using interaction terms of the form AGE
(1- d10)LAGE, where LAGE is the legal age of school entry or school exit.''
15. This is a more parsimonious specification than the one with all possible interaction terms. In
particular, separate coefficientson the dummy variable (1- d10)and their interactions with age, gender,
and rural residence did not add to the explanatory power of the child labor equation.
44 T H E W O R L D BANK E C O N O M I C REVIEW, VOL. 20, N O . I
The child's value of time in school will depend on how much the child can
learn. This will depend on home attributes that are complementary to child time
in school, such as books and parental education, and on the quality of the
school. Most of these measures are self-explanatory. However, some of the
school variables merit comment. The measure of the classroom environment,
inadequacy, is a weighted average of several measures of poor school infrastruc-
ture and supplies. Teachers were asked the extent to which they judged class-
room lighting, temperature, hygiene, security, acoustics, and textbooks to be
inadequate. The weighted sum of the responses is used as the aggregate index of
school shortcomings, where the weights were taken as the first principal com-
ponent from a factor analysis of the teachers' responses. The number of Spanish-
or Portuguese-speaking students is included as a measure of the cost of providing
schooling services. As the number of nonnative speakers of the language of
instruction increases, resources must be diverted to second-language instruction,
potentially limiting school productivity.
IV. ECONOMETRIC STRATEGY
The results in table 1 suggest a strong negative effect of child market labor on
school achievement, but the effect may be in the reverse direction-poorschool-
ing outcomes leading to child labor. The direction of this bias is difficult to
predict. The most plausible is that poor school performers are sent to work so
that the least squares coefficient on child labor will be biased downward.
However, both Tyler (2003)and Stinebrickner and Stinebrickner (2003)found
biases in the opposite direction for older students, with better students more
likely to work. Measurement error in the self-reported incidence of child labor
could also bias the estimated coefficient of child labor on schooling outcomes.
The cumulative direction of these sources of bias cannot be established, but both
simultaneity and measurement error can be handled by the use of plausible
instruments that alter the probability of engaging in child labor without directly
affecting test scores.
The first step in the estimation process is to predict child labor. The cate-
gorical measure of child market work includes 0 (almostnever work),1 (some-
times work), and 2 (often work). Equation (1)was estimated with an ordered
probit specification, using child, parent, school, and community variables to
explain variation in market work. Predicted child labor from equation (1)is
used as the measure of C in estimating equation (2).This two-stage estimation
leads to consistent, but inefficient estimates of the parameters of the achieve-
ment equation. A bootstrapping method is used to correct for the inefficiency
in the estimators in which 100 samples with replacement are drawn from the
original data, subjected to the ordered probit estimation and then inserted into
the second-stage achievement equation to simulate the sampling variation in
the estimates. The bootstrap standard errors are reported for the test score
equations.
Gunnarsson, Orazem, and Sanchez 45
Estimates from the probit child labor supply equation, reported in table 2, are
needed to identify the effect of child labor on test scores but are also of interest in
their own right. The estimation uses the dependent variables reported in table 1
except that data for Mexico and Venezuela are dropped because child's age was
not reported. Because the two samples are not identical, separate estimates are
reported for the samples of children taking the mathematics and language
examinations. The coefficients on the age-interacted variables differ somewhat
across the two samples, but the overall relationship between age and child labor
is similar between the two samples. The other coefficient estimates are similar
across the two samples.
Boys are more likely than girls to work outside the home, and rural boys and
girls work more than their urban counterparts, who in turn work more than
their metropolitan counterparts. Children of more-educated parents and
TABLE 2. Ordered Probit Regression Results on Child Labor
Variable Mathematics Test Scores Language Test Scores
Exogenous variables
Child
Age 0.048 (0.009)** -0.014(0.009)
BOY 0.291 (0.036)"" 0.163 (0.037)**
No preschool -0.016 (0.019) 0.029 (0.019)
Parents/household
Parent education -0.065(0.007)"' -0.046(0.008)**
Books at home -0.080 (0.012)'= -0.071(0.012)'"
School
Spanish enrollment/lOO -0.004 (0.002)"" -0.005 (0.002)~"
Inadequate supply 0.062 (0.009)*" 0.065 (0.009)+*
MatNweek (SpanisNweek) -0.014 (0.004)"* -0.010 (0.003)*+
Community
Rural 0.350 (0.033)'" 0.290 (0.034)"'
Urban 0.197 (0.033)"" 0.121 (0.031)""
Instruments
Boyxrural -0.019 (0.045) 0.144 (0.045)""
Boyxurban -0.062(0.043) 0.103 (0.044)+"
Agexcompulsory start (1-d10) 0.004 (0.001)"" 0.002 (0.001)"
Agexcompulsory end (1-dl 0) -0.002 (0.000)"" 0.000 (0.001)
LL -21,623.743 -21,179.099
Pseudo R~ 0.034 0.034
Number of observations 20,699 20,290
"Significant at the 0.10 confidence level.
"+Significantat the 0.05 confidence level.
Note: Numbers in parentheses are standard errors. Regressions also include dummy variables
that control for missing values.
Source: Authors' computations based on data from the 1997 survey by the Latin-American
Laboratory of Quality of Education, as described in the text; UNESCO (2002).
children who have access to more books in the home are less likely to work
outside the home. School quality also affects the incidence of child labor. Schools
with inadequate supplies encourage child labor. Children in schools with more
non-Spanish or non-Portuguese language speakers among their peers are also
more likely to work outside the home.Schools that offer more classes in Spanish
or Portuguese and mathematics per week also lower the incidence of child labor.
In general, these results suggest that better schooling inputs in the home and
school lower the incidence of child labor. The exception is that attending pre-
school does not have a significant effect on child labor in this sample.
The joint test of the null hypothesis that the instrumental variables have no
effect on child labor is easily rejected. Variation in truancy laws across countries
and in the child labor market for boys within countries does shift the probability
that children work. The impact of these laws on the average incidence of child
labor is illustrated in figures 1 and 2. The effect was disabled below age 10. As
the school starting age rises from ages 5 to 7, the probability of child labor rises
about 6 percentage points for a 10-year-old, all else equal, and by 10 percentage
points for a 14-year-old (figure1).As the school-leavingage rises from 12 to 16
years old, the probability of child labor falls by 8.5 percentage points for a
10-year-old and by 11.5 percentage points for a 14-year-old (figure 2). These
results suggest that truancy laws do have an effect on child labor on average.
Regional variation in the market for child labor shifts child labor supply for
boys and girls (figure 3). The dummy variable spline effectively fixes child labor
intensity for children under 10 years of age. After the age of 10, child labor
intensity rises for both boys and girls. In each market, boys work more than
girls.16The higher market labor force participation for boys is consistent with the
presumption that the marginal product of child labor is higher for boys than girls.
However, rural girls have higher labor force participation than metropolitan boys.
VI. CHILD LABOR AND SCHOOL ACHIEVEMENT
The results from estimating equation (2)both with and without controls for the
endogeneity of child labor are reported in table 3. In the specificationin table 3,
when child labor is treated as exogenous, it takes the values of 0 (almost never
work), 1 (sometimes work), or 2 (often work). When treated as endogenous,
child labor is a continuous variable with domain over the real line taken as the
fitted values from the ordered probit estimation in table 2. The rest of the
regressors are the child, household, parent, and school variables used as regres-
sors in table 2.17
16. Ages are truncated below 8 (0.4percent of the sample) and above 15 (0.8 percent of the sample)
because of insufficient observations to generate reliable child labor supply trajectories.
17. Similar estimates of the adverseeffect of child labor on test scoreswere obtained when a school-
specific fixed effect was used to control for the impact of variation in school and community variables
instead of the vector of school and community variables.
TABLE 3 . Least Squares and Instrumental Variables Equations on Test Scores
Child Labor Exogenousa Child Labor ~ n d o ~ e n o u s ~
Variable Mathematics Test Scores Language Test Scores Mathematics Test Scores Language Test Scores
Work outside -1.184(0.051)"" -1.087(0.036)'" -7.603(1.248)'" -3.980(0.484)""
Beta coefficientc [-0.1591 [-0.2041 [-0.4081 [-0.2951
Child
Age 0.097 (0.027)'" 0.045 (0.019)*" 0.309 (0.070)"" 0.162 (0.024)""
BOY 0.731 (0.079)"" -0.165(0.056)"" 2.480 (0.358)"" 0.679 (0.155)""
No preschool -0.256 (0.093)"" -0.181(0.066)"" -0.376(0.088)'" -0.079(0.040)'"
Parents/household
I'arent education 0.327 (0.036)"" 0.280 (0.026)'" -0.107(0.106) 0.134 (0.042)'"'
Books at home 0.735 (0.061)"" 0.497 (0.042)"" 0.196 (0.100)"" 0.258 (0.037)""
School
Spanish enrollmentl100 -0.046(0.008)"" 0.022 (0.006)"" -0.079(0.010)"" 0.007 (0.005)
Inadcquate supply -0.329(0.046)"" -0.357(0.031)"" 0.073 (0.096) -0.140(0.038)""
Mathiweek (Spanishiweek) 0.027 (0.017) 0.022 (0.006)"" -0.073 (0.016)"" -0.049(0.012)""
Community
Urban 0.730 (0.107)"" 0.240 (0.076)'" 1.847 (0.225)"" 0.794 (0.117)""
Rural -0.692(0.122)"" -0.893(0.087)"" 1.641 (0.410)"" 0.275 (0.202)
Constant 13.778 (0.446)"" 10.657 (0.248)*" 14.400 (0.453)"* 8.045 (0.391)""
R~ 0.084 0.127 0.063 0.091
Number of observations 20,699 20,290 20,699 20,290
"Significantat the 0.10 confidence level.
"Significantatthe0.05confidencelevel.
Note: Regressionsalso include dummy variables controlling for missing values.
aNumbersin parentheses are standard errors.
b~umbersin parentheses are bootstrap standard errors.
'The beta coefficients indicate the number of standard deviations the test score will change from a 1 standard deviation increase in child labor.
Source:Authors' computations based on data from the 1997 survey by the Latin-American Laboratory of Quality of Education, as described in the text;
UNESCO (2002).
The impact of child labor on test scoresis negativeand significantwhether child
labor is treated as exogenous or endogenous.18Because of the difference in the
scaleof measured child labor across the two specifications, it is difficult to directly
compare the magnitude of the implied effect of child labor on test scores. The
results are compared in two ways. First, the implied effect of a 1 standard
deviation increase above the mean in child labor is computed in each of the
equations. When treated as exogenous, a 1 standard deviation increase in child
labor causes both mathematics and language tests scores to fall by about 0.2
standard deviations. In other words, childrenworking 1 standard deviation above
the mean score on average 8 percent lower on mathematics examinations and
6 percent lower on language examinations than do otherwise identical children
working at the mean level. When controllingfor endogeneity, the effect increases
to 0.4 standard deviation (16percent)drop in the mathematicsexamination and a
0.3 standard deviation (11 percent) drop in the language examination. This
finding that the magnitude of the child labor effect on academic achievement
rises after controlling for endogeneity is consistent with results reported by Tyler
(2003)and Stinebrickner and Stinebrickner (2003)for older U.S. students.
Second, the two sets of estimates are compared by tracing the predicted
mathematics and language test scores at each decile of the reported and pre-
dicted child labor distributions (figures 4 and 5). At the breakpoints of the
exogenous measure (going from child labor level 0 to level 1 at the 40th
percentile and from level 1 to level 2 at the 74th percentile), the predicted test
scores using the reported and corrected measures are close to one another.
However, the relationship is steeper at the upper and lower tails of the distribu-
tion of predicted child labor, particularly for the mathematics test. The implica-
tion is that the impact of child labor on test scores is understated in the first two
columns of table 3 by restricting the range of child labor to three discrete levels.
Glewwe's (2002)review of the human capital literature in developing coun-
rries argued that cognitive ability as measured by test scores is strongly tied to
later earnings as an adult. Returns to schoolingfor those who worked as children
would therefore be expected to be lower than for those who did not work, all
else equal. Consistent with that expectation, Ilahi, Orazem, and Sedlacek (forth-
coming) found that, holding constant years of schooling completed, Brazilian
adults who worked as children received 4-11 percent lower returns per year of
schooling completed. The estimates here suggest that child labor outside the
home reduces achievement per year of schooling attended by 11-16 percent.
Because many of the third and fourth graders in the sample will repeat the grade,
the estimates are an upper-bound measure of the lost human capital per year
18. The Davidson-MacKinnon (1993, pp. 237-40)variant of the Hausman test easily rejected the
assumption of exogeneity of child labor. The overidentification tests of the instruments failed to reject the
null hypothesis of ekogeneity at the 10th percentile in the language test sample and at the 5th percentile
for the mathematics test sample.
Gunnarsson, Orazem, and Scinchez 49
FIG RE 4. Predicted Language Test Scores by Child Labor Decile
U
I
-.--.
--.---..-...._. I
-..---___._ I
--..
--~~---.-._.__..-
--....
Orderd probit estimate, chlld --....
--...
..-._.
labor endogenous --.
labor exogenous
.-..-..
-......
- - _ _
-..... i
I.-__
....--.
--..._
---..
....
I -..
-..-...
-.....-..
.-.
Maximum= 19, Averagescore = 11.3
6 L 1
0 10 20 30 0 50 60 70 80 M 100
Childlabor decile
Note: Dashed lines shows 1 standard deviation confidence band for ordered probit estimates.
Source: Authors' simulations based on results in table 3, column 4.
completed, and so the results correspond closely in magnitude to the estimates of
Ilahi, Orazem, and Sedlacek of adverse impacts of child labor on earnings.
Most of the other variables have similar effects across the two sets of estimates
in table 3, with two main exceptions. The adverseeffects of being a boy or being
in a rural school disappear in the instrumented equations. Gender and rural
residence are closely tied to the incidence of child labor. It is likely that the
negative effects on test scores of being male and being in a rural area are related
to the indirect effect of these variables on the higher probability that male and
rural children work.
Parental education and availability of books in the home lose influence on test
scores after controlling for the endogeneity in child labor. School attributes also
become less important in explaining test scores. Again, these factors had strong
negative effects on child labor, and so part of their positive effect on school
outcomes presumably works through their impact on child school attendance
and reduced time at work. The literature on the extent to which school quality
can explain variation in school achievement has emphasized the large variation in
coefficients for the same school inputs across studies and country settings
(Hanushek and Luque 2003). The results here suggest that one reason for the
uncertainimpact of school attributes may be that school qualityis more important
in affecting school attendance and child labor than in directly affectingtest scores.
50 THE W O R L D BANK E C O N O M I C REVIEW, VOL. 20, N O . I
FIGURE 5 . Predicted Mathematics Test Scores by Child Labor Decile
I Maximum =32, Average sras = 14.6
0 10 20 30 40 91 50 70 80 �I 101
Childlabor decile
Note: Dashed lines shows 1 standard deviation confidence band for ordered probit estimates.
Source: Authors' simulations based on results in table 3, column 3.
VII. CONCLUSIONS
Working outside the home lowers average school achievement in samples of third
and fourth graders in each of the nine Latin-American countries studied. Child
labor is shown to have significant adverse effects on mathematics and language
test scores using various specifications correcting for possible endogeneity and
measurement error in self-reported child labor intensity. Children who work even
occasionally score an average of 7 percent lower on language examinations and
7.5 percent lower on mathematics examinations. There is some evidence that
working more intensely lowers achievement more, but these results are more
speculative in that empirical models were unable to distinguish clearly between
working "sometimes" and working "often."
These adverse effects of child labor on cognitive ability are consistent in magni-
tude with the estimated adverseeffects of child labor on earnings as an adult. Thus,
it is plausible that child labor serves as a mechanism for the intergenerational
transmission of poverty, consistent with empirical evidence presented by Emerson
and Souza (2003)and the theoretical models of poverty traps advanced by Basu
(2000),Basu and Van (1998),and Baland and Robinson (2000).
Such large effects suggest that efforts to combat child labor may have substan-
tial payoffs in the form of increased future earnings or lower poverty rates once
children become adults. How to combat child labor is less clear. The child labor
Gunnarsson, Orazern, and Sanchez 51
supply equations developed here suggest that truancy laws have some effect in
lowering the incidence of child labor. However, most of the variation in child
labor occurs within countries and not across countries, so policies must address
local child labor market and poverty conditionsas well as national circumstances
in combating child labor. Policies that alter the attractiveness of child labor or
bolster household income, such as income transfer programs that condition
receipt on child enrollment or reduced child labor, are likely canddates. Recent
experience with such programs in Brazil, Honduras, Mexico, and Nicaragua
appears to support further development and expansion of such programs.
TABLE A- 1. Variable Description
Endogenous variables
Math score Mathematics test score (C)
Language score Language test score (C)
Work outside Index of how often student works outside the home (0-2)(C)
Often Student reports that he or she often works outside the home (C)
Sometimes Student reports that he or she sometimes works outside the home (C)
Almost never Student reports that he or she almost never works outside the home (C)
Exogenous variables
Child
Age Student age (years)(C)
dl0 Dummy variable if student is below 10 years old
BOY Dummy variable if student is a boy (C)
No preschool Student did not attend preschoolkindergarten (C)
Parents/household
Parent education Average education of parent(s) or guardian(s)(P)
Books at home Number of books in student's home (P)
School
Spanish enrollment Total number of Spanish (Portuguese) speaking students enrolled (Pr)
Inadequate supply Index of school supply inadequacy (Pr)
Mathiweek Number of mathematics classes per week (Pr)
SpanisMweek Number of Spanish (Portuguese)classes per week (Pr)
Community (Reference:
Metropolitan area with 1 million people or more)
Urban Dummy variable indicating if school is located in an
urban area (2,500 to 1 million people) (S)
Rural Dummy variable indicating if school is located in a rural area
(fewer than 2,500 people) (S)
Instruments
Legal structure
Compulsory start Compulsory school starting age in the country (U)
Compulsory end Compulsory school ending age in the country (U)
Note: C, child survey or test; P,parent's survey;T, teacher's survey; Pr, principal's survey;S, survey
designer's observation of socioeconomic characteristicsof school community; UNESCO estimate.
Souvce: Authors' analysis based on data from the 1997 survey by the Latin-American Labora-
tory of Quality of Education, as described in the text; UNESCO (2002).
TA E A-2. Summary Statistics
BL
Number of Standard
Variable Observations Mean Deviation Minimum Maximum
Endogenous variables
Mathematics score
Language score
Work outside
Often
Sometimes
Almost never
Exogenous variables
Child
Age
d l 0
BOY
No preschool
Parents/household
Parent education
Books at home
School
Spanish enrollment
Inadequate supply
MatWweek
Community
Urban
Rural
Instruments
Compulsory start
Compulsory end
Source: Authors' computations based on data from the 1997 survey by the Latin-American
Laboratory of Quality of Education, as described in the text; UNESCO (2002).
Akabayashi, Hideo, and George Psacharopoulos. 1999. "The Trade-Off Bemeen Child Labor and
Human Capital Formation: A Tanzanian Case Study." The Journal of Development Studies 35(5):
120-40.
Angrist, Joshua D., and Alan B. Krueger. 1991. "Does CompulsorySchool Attendance Affect Schooling
and Earnings?"The Quarterly Journal of Economics 106(4):979-1014.
Baland, Jean-Marie, and James A. Robinson. 2000. "Is Child Labor Inefficient?"TheJournal of Political
Economy 108(4):663-79.
Basu, Kaushik. 2000. Analytical Development Economics: The Less Developed Economy Revisited.
Cambridge, MA: MIT Press.
Basu, Kaushik,and Pham Honang Van. 1998. "The Economicsof Child Labor." The American Economic
Review 88(3):412-27.
Gunnarsson, Orazem, and Sanchez 53
Basu, Kaushik, and Zafiris Tzannatos. 2003. "The Global Child Labor Problem: What Do We Know and
What Can We Do?" The World Bank Economic Review 17(2):147-73.
Ben Porath, Yoram. 1967. "The Production of Human Capital and the Lifecycleof Earnings." TheJournal
of Political Economy 75(4):352-65.
Card, David. 1999. "The Causal Effect of Education on Earnings." In Orley Ashenfelter and David Card,
eds., Handbook of Labor Economics, vol. 3A. Amsterdam: Elscvier.
D'Amico, Ronald. 1984. "Does Employment During High School Impair Academic Progress?"Sociology
of Education 57(3):152-64.
Davidson, Russell, and James G. MacKinnon. 1993. Estimation and Inference in Econometrics. Oxford:
Oxford University Press.
Edmonds, Eric V., and Nina Pavcnik. 2005. "Child Labor in the Global Economy." The Journal of
Economic Perspectives 19(1): 199-220.
Ehrenberg, Ronald G., and Daniel R. Sherman. 1987. "Employment While in College, Academic Achieve-
ment, and Postcollege Outcomes." TheJournal of Human Resources 22(1):1-23.
Emerson, Patrick M., and AndrC Portela Souza. 2003. ''Is There a Child Labor Trap? Inter-Generational
Persistence of Child Labor in Brazil." Economic Development and Cultural Change 51(2):375-98.
Glewwe, Paul. 1996. "The Relevance of Standard Estimates of Rates of Return to Schooling for Educa-
tional Policy: A Critical Assessment."Journal of Development Economics 51(2):267-90.
-. 2002. "Schools and Skills in Developing Countries: Educational Policies and Socioeconomic
Outcomes." Journal of Economic Literature 40(2):436-82.
Hanushek, Eric A. 1995. "Interpreting Recent Kesearch on Schooling in Developing Countries." The
World Bank Research Observer lO(2):22746.
Hanushek, Eric A., and Javier A. Luque. 2003. "Efficiency and Equity in Schoojs Around the World."
Economics of Education Review 22(5):481-502.
Heady, Christopher. 2003. "What is the Effectof Child Labour on Learning Achievement?Evidence from
Ghana." World Development 31(2):385-98.
Heckman, James J. 1976. "A Life-Cycle Model of Earnings, Learning, and Consumption." Journal of
Political Economy 84(4):S11-44.
Howard, Ian. 1998. "Does Part-Time Employment Affect A-Level Grades Achieved?" PSSI Forum N26
(October):10-11.
Huffman, Wallace E., and Peter F. Orazem. 2006. "Agriculture and Human Capital in Economic Growth:
Farmers, Schooling, and Nutrition." In R. E. Evenson, P. Pingali, and T. P. Schultz, eds., Handbook of
Agricultural Economics, vol. 3. Amsterdam: North Holland.
Ilahi, Nadeem, Peter F. Orazem, and Guilherme Sedlacek. Forthconling. "Child Labor, Schooling, and
Poverty in Latin America." In Peter F. Orazem, Guilherme Sedlacek, and P. Zafiris Tzannatos, eds.,
Child Labor and Education in Latin America. Washington, D.C.: IntcrAmerican Development Bank.
Kremer, Michael R. 1995. "Research on Schooling: What We Know and What We Don't; A Comment on
Hanushek." The World Bank Research Observer lO(2):247-54.
Levy, Victor. 1985. "Cropping Pattern, Mechanization, Child Labor, and Fertility Behavior in a Farming
Economy: Rural Egypt." Economic Development and Cultural Change 33(4):777-91.
Lillydahl, Jane H. 1990. "Academic Achievement and Part-Time Employment of High School Students."
Journal of Economic Education 21(3):307-16.
Moll, Peter. 1998. "Primary Schooling, Cognitive Skills and Wages in South Africa." Economics 65(258):
263-84.
Patrinos, Harry A., and George Psacharopoulos. 1997. "Family Size, Schooling, and Child Labor in
Peru-An Empirical Analysis."Journal of Population Economics lO(4):387-405.
Post, David, and Suet-ling Pong. 2000. "Employment During Middle School: The Effects on Academic
Achievement in the U.S. and Abroad." Educational Evaluation and Policy Analysis 22(3):273-98.
5 4 THE W O R L D BANK E C O N O M I C REVIEW, VOL. 20, N O . I
Psacharopoulos, George. 1997. "Child Labor versus Educational Attainment: Some Evidencefrom Latin
America." Journal of Poptrlation Economics lO(4):337-86.
Ray, Ranjan. 2000. "Child Labor, Child Schooling, and Their Interaction with Adult Labor: Empirical
Evidence for Peru and Pakistan." The World Bank Economic Review 14(2):347-67.
Ravallion, Martin, and Quentin Wodon. 2000. "Does Child Labor Displace Schooling? Evidence on
Behavioral Responses to an Enrollment Subsidy." The Economic Joural llO(462):(358-75.
Rosati, Furio Camillo, and Mariacristina Rossi. 2003. "Children's Working Hours and School Enroll-
ment: Evidence from Pakistan and Nicaragua." The World Bank Economic Review 17(2):283-95.
Rosenzweig, Mark R. 1980. "Neoclassical Theory and the Optimizing Peasant: An Econometric Analysis
of Market Family Labor Supply in a Developing Country." The Quarterly Journal of Economics 94(1):
31-55.
Rosenzweig,Mark R., and Robert Evenson.1977. "Fertility, Schoolingand the Economic Contribution of
Children in Rural India: An Econometric Analysis." Econometrics 45(5):1065-79.
Sedlacek, Guilherme, Susanne Duryea, Nadeem Ilahi, and Masaru Sasaki. Forthcoming. "Child Labor,
Schooling, and Poverty in Latin America." In Peter F. Orazem, Guilherme Sedlacek, and P. Zafiris
Tzannatos, eds., Child Labor and Education in Latin America. Washington, D.C.: InterAmerican
Development Bank.
Singh, Kusum. 1998. "Part-Time Employment in High School and Its Effect on Academic Achievement."
The Journal of Educational Research 91(3):131-39.
Singh, Kusum, and Mehmet Ozturk. 2000. "Effect of Part-Time Work on High School Mathematics and
Science Course Taking." The Joumal of Educational Research 91(2):67-74.
Stern, David. 1997. "Learning and Earning: The Value of Working for Urban Students." ERIC Digest
128. New York: ERIC Clearinghouse on Urban Education, Teachers College, Columbia University.
Stinebrickner, Ralph, and Todd R. Stinebrickner. 2003. "Working During School and Academic Perfor-
mance." Journal of Labor Economics 21(2):449-72.
Tyler, John H. 2003. "Using State Child Labor Laws to Identify the Effect of School-to-Work on High
School Achievement."Journal of Labor Economics 21(4):381-408.
UNESCO (United Nations Educational, Scientific, and Cultural Organization). 2002. Education Statistics
2001-RegionalReport on Latin America. Paris. (http://portal.unesco.org/uis).
The Long-Run Economic Costs of AIDS: A Model
with an Application to South Africa
Clive Bell, Shantayanan Devarajan, and Hans Cersbach
Primarily a diseaseof young adults, l u ~imposes economic costs that could be devastat-
s
ingly high in the long run by undermining the transmission of human capital-the main
driver of long-run economic growth-across generations. AIDS makes it harder for
victims' children to obtain an education and deprives them of the love, nurturing, and
life skills that parents provide. These children will in turn find it difficult to educate
their children, and so on. An overlapping generations model is used to show that an
otherwise growing economy could decline to a low-level subsistence equilibrium if hit
with an ms-type increasein premature adult mortality. Calibrating the model for South
Africa, where the HIV prevalence rate is over 20 percent, simulations reveal that the
economy could shrink to half its current size in about four generations in the absence of
intervention. Programs to combat the disease and to support needy families could avert
such a collapse, but they imply a fiscal burden of about 4 percent of GDP.
While the costs of AIDS in terms of human sufferingare undeniably large, estimates
of the associated macroeconomic costs have tended to be more modest, whether
their basis be an explicitly formulated economic-demographc model or cross-
country regression analysis. Most earlier studies of the former kind that focus on
Africa-the continent where the epidemic has hit the hardest-put the annual loss
of GDP at about 1percent.1These estimates stem from a particular view of how the
economy functions: the ms-induced increase in mortality, even if it reduceslabor
supply, also reduces the pressure of population on existing land and capital,
thereby raising the productivity of labor. Even if there is an accompanying decline
Clive Bell is a professor of economics at the South Asia Institute of the University of Heidelberg; his
email address is clive.bell@urz.uni-heidelberg.de. Shantayanan Devarajan is the chief economist of the
South Asia Region at the World Bank; his cmail address is sdevarajan@worldbank.org.Hans Gersbach is
a professor of economics at the Alfred-Weber-Institut of the University of Heidelberg; his emailaddress is
gersbach@uni-heidelberg.de. The authors thank Ramona Bmhns for excellent programming and research
assistance. They are also grateful to Bernhard Pachl, Lars Siemers, and participants in seminars at the
Universityof California at Berkeley, Cornell University, the Center for Global Development, the London
School of Economics, and the World Bank, as well as to three anonymous referees and the editor of this
journal, for their valuable and constructive comments and suggestions. The authors are responsible for
any remaining errors of analysis.
1. See, for example, Arndt and Lewis (2000);Cuddington (1993);Cuddington and Hancock (1994);
Kambou, Devarajan, and Over (1992);and Over (1992).
mWORLD BANK ECONOMIC REVIEW, VOL. 20, NO. 1, pp. 55-89 doi:l0.1O93/wberllhj006
Advance Access publication April 11, 2006
O The Author 2006. Published by Oxford University Press on behalf of the International
Bankfor Reconstructionand Development/ m worn BANK. All rightsreserved. For permissions,
please e-mail: joumals.permissions@oxfordjournals.org.
in aggregatesavings and investment (from the reallocation of expenditures toward
medical care, for instance), the net impact on the growth of GDP per capita turns
out to be small. Econometric investigations based on country panel data yield the
same result. Bloom and Mahal(1997),for example, found no effect on GDP at all;
on returning to the question later with new data, Bloom and others (2001)
managed to extract a small adverse effect.
This article argues that the long-run economic costs of AIDS are almost certain
to be much higher-and possibly devastating. In doing so, it joins company with
some other authors (Corrigan, Glomm, and Mindez 2004, 2005; Ferreira and
Pessoa 2003),who recently and independently have pursued an approach based
on an overlapping generations framework. This approach involves a very differ-
ent view of how the economy functions over the long run, one that emphasizes
the importance of human capital and its transmission across generations. The
accumulation of human capital-that is, the stock of knowledge and abilities
embodied in the population-is the force that generates economic growth over
the long run. The mechanism that drives the process is the transmission of
knowledge and abilities from one generation to the next.
The implications of this model are particularly relevant to Africa, the con-
tinent with the lowest level of human capital and the highest prevalence of the
disease.In many African countries, AIDS presents a formidable hurdle to long-run
economic growth. The application of the model to South Africa, that Sub-
Saharan outlier with relatively high levels of income and human capital (and
HIV prevalence), reveals that in the absence of specific interventions, a decline
from middle-income status is possible in the long run.
The argument establishing how AIDS can severely retard economic growth is
made in three steps.2 First, AIDS destroys existing human capital in a selective
way, striking primarily young adults. Some years after they have been infected, it
reduces their productivity by making them sick and weak. It then kills them in
their prime, destroying the human capital formed in them through child-rearing,
formal education, and learning on the job.
Second, AIDS weakens or even wrecks the mechanisms that create human
capital in the next generation. The quality of child-rearing depends heavily on
the parents' human capital. If one or both parents die before their offspringreach
adulthood, the transmission of knowledge and potential ~roductivecapacity
across the two generations will be weakened. At the same time, the loss of
income due to disabilityand early death reduces the lifetime resources available
to the family, which can lead to the children spending much less time, if any, at
school. The chance that the children will contract the disease in adulthood also
makes investment in their education less attractive, even when both parents
themselves remain uninfected. The weakening of these transmission processes
2. The argument here is confined to those factors that are the most important. For a longer list of the
epidemic's economic effects and related discussion, see, for example, Bell, Devarajan, and Gersbach
(2004)and Corrigan, Glomm, and Mendez (2005).
Bell, Devarajan, and Gersbach 57
is insidious: its effects are felt only over the longer run, as the poor education of
children today translates into low productivity of adults a generation hence.
Third, as the children of AIDS victims become adults with little education and
limited knowledge receivedfrom their parents, they are less able to invest in their
own children's education, and a vicious cycle ensues. If nothing is done, the
outbreak of the disease can eventually precipitate a collapse of economic pro-
ductivity. Early in the epidemic, the damage may appear to be slight, but as the
transmission of capacities and potential from one generation to the next is
progressively weakened and the failure to accumulate human capital becomes
more pronounced, the economy will begin to slow down, with the growing
threat of a collapse to follow.
The argument has two important implications for economic policy. The first
is fiscal. By killing off mainly young adults, AIDS also seriously weakens the tax
base and thus reduces the resources available to meet the demands for public
expenditures, including those aimed at accumulating human capital, such as
education and health services not related to NDS. As a result, the state's finances
will come under increasing pressure, exacerbated by the growing expenditures
on treating the sick and caring for orphans.
The other effect is an increase in inequality. If orphaned children are not given
the care and education enjoyed by those whose parents remain uninfected, the
weakening of the intergenerational transmission mechanism will express itself in
increasing inequality among the next generation of adults and the families they
form. Social customs of adoption and fostering, however well established, may
not be able to cope with the scale of the problem, thereby shifting the onus onto
the government, which is likely to experience increasing fiscal difficulties and
thus to lack the resources to assume this additional burden.
The policy objective, therefore, is to avoid such a collapse. The instruments
available for this purpose are (a)spending on measures to contain the disease
and treat the infected, (b) aiding orphans, in the form of income support or
subsidies contingent on school attendance, and (c)taxes to finance the expendi-
ture program. The central policy problem is to find the right balance among
these interventions to ensure economic growth over the long run without exces-
sive inequality.
This article relates to recent contributions to the literature as follows. Those
that adopt an overlapping generations framework have chosen somewhat dif-
ferent points of emphasis. In the model here, higher mortality risk undermines
the formation of human capital through three channels. First, if one or both
parents die early, their children will have less productive capacity because less
human capital is transmitted. Second, the loss of income caused by early death in
a family reduces schooling. Third, the chance that the children will be infected as
adults makes investment in their education less attractive. Corrigan, Glomm,
and Mindez (2004, 2005) consider only the first two channels, but they allow
for effectson the accumulation of physical capital, which are absent in the model
here. Thus, this article complements theirs in the task of establishing how AIDS
might influencethe course of per capita income. As for the possible magnitude of
these effects, Corrigan, Glomm, and Mkndez (2004) calibrate their model to
Sub-Saharan economies and find that for infection rates3 of around 15-20
percent, the growth rate of per capita income drops about 30-40 percent.
Ferreira and Pessoa (2003)concentrate on the reduced returns to investment in
schoolingin a setting free of uncertainty, and estimate that the time devoted to it
can decline by up to a half.
Young (2005)adopts a quite different perspective on how the AIDS epidemic
impinges on the South African economy. He embeds a Beckerian household
model, with endogenous participation, fertility, and education decisions, in a
Solovian constant-savings-rate macroeconomic framework. In estimating the
behavioral equations and simulatingthe evolution of the South African economy,
two competing effects are emphasized. On the one hand, the epidemic is likely to
have a negative impact on orphans' accumulation of human capital. On the other
hand, high prevalence rates lower fertility. Young finds that even with the most
pessimistic assumptionsregardingeducational attainment, the fertility effect dom-
inates and future per capita consumption possibilities are enhanced. Although
more channels through which the epidemic may harm human capital accumula-
tion are considered here, fertility is exogenous. Sensitivity tests are conducted,
however, and these reveal that changes in the level of fertility have only minor
effects on the growth of productivity. Bruhns (2005)develops a closely related
theoretical model in which households choose the level of fertility and applies it to
Kenya. Her conclusions are broadly similar to the ones arrived at here.
Some econometric studies look at aspects of the link between AIDSand human
capital. McDonald and Roberts (2004)estimate an augmented Solowmodel that
incorporates both health and education capital. They employ a panel of 122
countries over a longer timespan than that of Bloom and others (2001) and
conclude that the macroeconomic effects of HIVIAIDShave been substantial,
especially in Africa, where the average marginal impact on income per capita
of a 1 percent increase in the HIV prevalence rate is estimated to be -0.59
percent. Hamoudi and Birdsall (2004) provide indirect econometric evidence
that AIDS reduces schooling in Africa. Using data from Demographic and Health
Surveys conducted in 23 Sub-Saharan African countries and employing two
specifications, they settle on the estimate that a reduction in life expectancy at
birth of 10 years is associated with a fall of 0.6 years in the average schooling
attained by that cohort. Given that life expectancy at birth in most countries in
Southern and East Africa fell by at least 10 years over 1985-2000 (Dorrington
and Schneider 2001) and that average schooling among the population aged
25-49 years was in the modest range of 3-6 years, this is a significant and
disturbing finding. Although their measure of mortality differs from the one
used here, their finding supports the general approach adopted here. Other
3. The term is theirs, but a close reading strongly suggests that they mean prevalence rates.
Bell, Deuarajan, and Gersbach 59
recent microeconomic work suggests that orphans indeed suffer various set-
backs. Gertler, Levine, and Martinez' (2003)study of Indonesia11 children, for
example, shows that orphans are less healthy, less likely to go to school, and
overall less prepared for life. Case, Paxson, and Ableidinger (2002)found for a
group of Africancountries that the schooling of orphans depends heavily on how
closely related they are to the head of the adopting household.
In section I of this article, we tackle the question of how AIDS impinges on the
economy conceptually by extending the model of Bell and Gersbach (2001)to
deal with disease-ridden environments, in which premature adult mortality is
increased by the outbreak of an epidemic. Parents have preferencesover current
consumption and the level of human capital attained by their children. The
decision about how much to invest in education is influenced by premature
adult mortality in two ways: the family's lifetime income depends on the adults'
health status and the expected payoff depends on the level of premature mortal-
ity among the children when they reach adulthood.
In section 11, we apply the model to South Africa. The choice of South Africa
as a test bed is a natural one on several grounds. First, the very nature of the
model demands that the available economic and demographic series be long and
fairly reliable if the base for calibration is to be solid. Second, South Africa is a
middle-income country that has experienced substantial growth over much of
the past half century. A collapse of the kind analyzed in section I, were it to
occur, would therefore mean that there is a long way to fall. Third, the epidemic
has progressed rapidly in South Africa, from a prevalence rate among the
population aged 15-49 years of about 1 percent in 1990 to just over 20 percent
in 2003 (UNAIDS 2004).
Finally, in section 111, we examine policies to avert the long-run economic
decline caused by AIDS. Interventions in the spheres of health and education are
examined. Finding the right balance between these two sets of measures is the
central policy problem, and the results in this section attempt to illuminate how
the balance should be struck. In any case, the sheer magnitude of the problem
indicates that additional public spending of the order of 3-4 percent of GDP may
be needed to contain the epidemic and ward off its worst effects.
There are two periods of life, childhood and adulthood. On becoming adults,
individuals form families and have children. When the children are very young,
they can neither work nor attend school. Since investment in education is
assumed to be the only form of investment, the family's full income is wholly
consumed in this phase. Only after this phase is over, do the adults learn whether
they will die prematurely-and thus leave their children as half or full orphans.
Early in each generation of adults, therefore, all nuclear families fall into one of
the following four categories: (a)both parents survive into old age, (b)the father
dies prematurely, (c) the mother dies prematurely, and (d) both parents die
prematurely. These states are denoted by st E St := {1,2,3,4). The subjective
probability that a family formed at the start of period t lands in category st is
denoted by ~ , ( s , ) Once their states have been revealed, families make their
. ~
allocative decisions accordingly, and the formation of human capital takes place.
What follows is a terse account of the main elements; the details are set out in
appendix A1.
Human capital is formed by a combination of child-rearing, whose quality
depends on the parents' combined human capital, &(st), and the child's formal
education, et, expressed as a fraction of school-going years. A child so reared in
period t attains the following level of human capital in period t +1:
where z(st) represents the strength with which capacity is transmitted across
generations, f(.)can be thought of as the "educational technology," and the
presence of the 1 in the upper branch grants this basic (normalized) level of
human capital to wholly uneducated adults. 5(11) is the level of human capital
attained by full orphans who grow up without care or education.
Let an individual's output be proportional to his or her level of human capital,
an assumption that is certainly plausible over the very long run. Then a house-
hold with nt children that finds itself in state st will have a well-defined level of
full income, which the adults can allocate between consumption and investment
in the children's education. The latter pays off in the form of each child's human
capital on reaching adulthood. The (surviving) parents' optimal level of such
investment, ey[At(st),st, depends on the level of full income, the relative
price of education, the strength of their altruism toward their children, and the
expected level of premature adult mortality in period t +1,
as they subjectively estimate it in period t. Substituting ey into equation 1 yields
z(st>f(ey (At(st),st, ~;+1))~t(st)1, + st = 1, 2, 3
(2) l t + l ={ a st = 4
Equation (2)describes a random dynamical system. Note that each child in any
given family state st attains the same At+l in adulthood with certainty, but he or
she can wind up in any of the states st+l E {1,2,3,4) after reaching adulthood
and forming a family in period t +1, and the succeeding branches proliferate in
the future. The attendant threat of growing inequality will occupy an important
4. The populationis assumed to belarge enough that this is also the fraction of all families in thatstate
after all premature adult deaths have occurred. Observe that these probabilities change over the course of
the epidemic (table 2).
Bell, Devarajan, and Gersbach 61
place in the analysis of policy interventions, but there is no space to go into the
dynamical properties of the system in any detail here. What follows is aimed
only at clarifying certain of their qualitative features.
It suffices, for this particular purpose, to look at what happens when there is
no premature adult mortality [.rrt+l (1)= 1 for all t],so that the only state that is
ever observed is st = 1. To derive the typical dynamics, it is assumed that
altruism is not operative when the adults are uneducated, that is, e! = 0 when
itissufficientlycloseto1.Itcanthen beshownthatthesystemhasatleast two
stationary states with respect to human capital if z(l)f(l)2;la +1 >lwa, where 2
is the lowest level of an adult's human capital such that a two-parent household
chooses full education for the children in such an environment (Bell and
Gersbach 2001).
The resulting phase diagram is illustrated in figure 1, where Ad(> 2) denotes
the smallest endowment of the adults' human capital such that they begin to
send their children to school and Aa(= 21') denotes the corresponding endow-
ment at which children finally enjoy full-timeschooling. As depicted, the system
has just two stationary states with respect to human capital. One is the state of
economic backwardness, defined as A = 2. This stable state is a poverty trap,
wherein all generations are at the lowest level of human capital. The other is an
unstable state (A,= A*Vt), in which the parents' human capital is such that they
FIGU E 1. Phase Diagram Without Premature Adult Mortality
R
choose a positive level of education for their children, who then attain A*/2 in
adulthood. To be precise, and recalling equation (2),A* satisfies
where ~ ~ (=1 1)for all t. Observe that, starting from any A >A*, unbounded
growth is possible if and only if 2z(l)f(l)21, and that the growth rate then
approaches 2z(l)f(1)-1 asymptotically.
These results reveal that the intrusion of premature adult mortality may affect
the system's dynamics not only by changing the probabilities of the states but
also by increasingthe values of Ad, A* and Aa for states 1,2, and 3, respectively,
and thereby increasing the range of human capital levels within which a pro-
gressive decline into backwardness will set in. This turn of events is now
examined in more detail.
Disease, Increasing Inequality, and Economic Collapse
The process by which the outbreak of an epidemic like AIDS may lead to
economic collapse can be described as follows. At the start of period t = 0, a
society of homogeneous, two-parent families, each with adult human capital
endowment 2A0,is suddenly assailed by a fatal disease. While the children are
still young, all adults learn whether they are infected with the disease, and the
survivors then choose the consumption-education bundle (ct(s0),ei(s )) for o
so = 1,2,3.How does the outbreak affect the subsequent development of the
society? Children who are left as unsupported orphans (so= 4) fall at once
into the poverty trap. Even if both parents survive but have been such orphans
in childhood, they cannot afford to send their children to school (as assumed
above), and their succeeding lineage remains there. To discover what happens
to the rest, the critical value function ;l*(s,K) is introduced for s E {1,2,3),
which is defined for stationary fertility and mortality, n, = n, 'dt and
K, = K, 'dt. In this setting, it is natural to assume perfect foresight, namely
= K t=K 'it.
where A*(1)= 2A"(1), A*(2)= A*(3) = A*(2) = A*(3), and K is a sufficient
statistic of premature adult mortality in the stationary state, in which, by
definition, all expectations are realized. A*(s,K) is the stationary-state level of
human capital associated with a particular state s, that is, in any pair of
generations, parent or parents and offspring share the same state. Equation
(3)states that if adults with human capital A* find themselves in family state s
and the mortality environment K, they will make choices for their children
such that the latter will attain the same level of human capital on reaching
adulthood.
Bell, Devarajan, and Gersbach 63
The critical value function has two key properties, which are established in
Bell, Devarajan, and Gersbach (2003):
The first property implies that a permanent increase in premature adult
mortality may cause a group that was earlier enjoying self-sustaining growth
to fall into the poverty trap. The second property implies that single-parent
familiesgenerally need higher individual levels of human capital than two-parent
ones to escape the trap, in which case an increase in premature adult mortality
also increases the share falling into the poverty trap by increasing the proportion
of one-parent families.
In the long run, if nothing is done to support full orphans and the children of
needy, one-parent households, the share of uneducated families will grow until,
in the limit, the whole population is in a state of economic backwardness. Not
only do some adults meet an early death but the whole society descends pro-
gressivelyinto the poverty trap. Two questions arise. First, what are the chances
that the AIDSepidemicwill so increase the level of premature adult mortality as to
precipitate a collapse?Second, what arrangements for support and insurance are
there to prevent such a collapse?These questions are addressed with reference to
South Africa in the next section.
This section falls into two parts. In the first, we cover the results of the calibra-
tion rather than the procedure itself, the details of which can bte found in Bell,
Devarajan, and Gersbach (2003).The robustness of the calibration is examined
using a sensitivity analysis of the critical value function. In the second part, we
develop three benchmark simulations of the model so calibrated.
Calibration and Sensitivity Analysis
Beginningwith the fundamental difference equation (I),the parameters z(s), the
functional form f(e), and the boundary value of d are needed. In view of the
highly nonlinear nature of the system and the limited information available, the
form f (e) =e is chosen. Since the unit time period of the model is a generation,
with two overlapping generations, it is defensible to set the span of each at 30
years.
Inspection of the series for South African GDP reveals that the period from
1960 to 1975 was one of fairly steady and appreciable growth. This early
subperiod is viewed as plausibleinitial basisfor assessinghow the post-apartheid
economy ought to be able to perform over the long haul. Denoting calendar
years by the subscript k and ignoring child labor, GDP in year k is
where Lk and Ak denote the size of the labor force and the average level of
efficiency in that year, respectively (table I), and the parameter a is the
productivity of a unit of human capital. Since the labor force series begins
in 1965, that year is the starting point for the calibration procedure. The series
for e, is quinquennial and takes the form of the average years of schooling
among the population aged 25 years and older-for example, 4.06 years in
1960. Defining full schooling as 10 years (ages 6-15 years inclusive) yields an
average value of e for those born between 1905 and 1935 of 0.406, which is
denoted by e!o.
Employing equation (1)recursively, together with the relation between a
family's earnings and its endowment of human capital and the series in
table 1, yields the estimates z = 0.818, a = 3,419, and = 2.696. The final
step is to shift the starting point to 1960. As pointed out in the introduction, the
AIDS prevalence rate rose from about 1 percent in 1990 to just over 20 percent a
decade later. This is a strong argument for choosing 1990 as the date of the
outbreak of the epidemicin South Africa, and hence1960 as the starting point in
the chosen 30-year framework. The interpolation from table 1 implies that A
grew at an annual rate of 0.58 percent between 1965 and 1990; thus
A60 =~~~/(1.00582.620. = )~
Two comments on these estimates are in order. First, the parameter a has the
dimension of 1995 U.S. dollars per efficiency unit of labor per year. According
to these estimates, therefore, a two-parent household in 1960 with two econom-
ically active adults and all the children attending school full-time would have
had a family income of a160or $17,915. In the event of a complete collapse that
left the entire population uneducated, the family's income would be just $6,840
TABLE 1. GDP, Schooling, and Labor Force, 1960-95
Year Yk(1995U.S. Dollar) eB YklLk
Idk
0.406 Not available Not available
0.410 7.42 x lo6 9,220
0.447 8.24 x lo6 10,990
0.453 9.25 x lo6 12,230
0.461 10.34 x lo6 12,320
0.495 11.93 x lo6 11,100
0.500 13.58 x lo6 10,650
Not available 15.29 x lo6 9,880
Not available 16.98 x lo6 10,130
Source:World Bank 2002; Aarro and Lee 1996.
Bell, Devarajan, and Gersbach 65
TABLE 2. Family State Probabilities Corresponding to Premature Adult
Mortality Rates
Note: ,q, denotes the probability that an individual will reach the age of s + x years, conditioned
on reaching the age of x years. State probabilities do not correspond to the actual years shown but
to the steady states associated with each disease environment ( D = 0'1).
Source: Authors' computations based on Dorrington and others (2001).
in the absence of child labor. Second, the estimate of z yields the value of the
intergenerational growth factor when children attend school full-time, namely
22 = 1.636. This corresponds to an annual growth rate of productivity of about
1.64 percent over the long run, which seems rather modest in light of the East
Asian experience, but quite in keeping with South Africa's recent performance.
The form of social organization has thus far remained conveniently in the
background, but now that preferences must be specified, a definite choice is
unavoidable. For much of the period in question, South Africa was quite rural,
so one can make the case that there was widespread pooling of orphaned
children, with all surviving parents caring for all children. This arrangement is
a salient feature of the benchmark cases to be analyzed below. Let preferences
over current consumption and the children's attained level of human capital on
reaching adulthood be logarithmic:
where the state st = 0 denotes pooling, and a representative pair of surviving
adults cares for nt/6, children, all of whom are valued and treated identically.
Given that the calibration is anchored to 1960, both 660 and households'
expectations in 1960 concerning the level of 690 are needed. The realized
value of 690 was 0.86. The great reductions in mortality in those three decades
benefited children far more than adults, however, so that it is defensible to set
the expected value of 690 at the actual value of 660. Finally, it is assumed that in
1960 a representative couple, unaware of and untouched by A~DSin any way,
chose the average years of schooling attained by the generation born between
1935 and 1965. This yields the value b =33.45.
To complete the array of economic parameters, estimates of p, the fraction of
an adult's consumption to which each child has a claim, and y,a child's human
capital when employed as a child laborer, are needed. Setting ,B at 0.5 seems
unobjectionable. A much lower value of yis called for: y= 0.2 yields a maximal
level of annual earnings from a child's labor of cry = $685, which may be on the
high side, but this is balanced by the fact that no direct costs of schooling have
been included.
Turning to the demographic components of the model, the population
roughly doubled between 1960 and 1990, so that in keeping with the assump-
tions in section I and the generation span of 30 years, each mother had, on
average, four surviving children over that period. Whether AIDS will affect
fertility in the future is unclear (some evidence points to a modest decline), but
what is certain is that AIDS has already contributed to a marked rise in mortality
among children under the age of 5 years (Dorrington and others 2001). Since
there is also some evidence that fertility had started to fall by the early 1990s
(World Bank 2002), it is assumed that each mother will have three surviving
children from 1990 onward.
The overriding concern in calibrating the model demographically, therefore, is
with premature mortality among adults. The benchmark caseis that where there is
no epidemic (D = O), which, in view of the low prevalence rate in 1990, is taken
to be the age-specific mortality profile for that year, as set out in Dorrington and
others (2001).The second reference case is that where the epidemic has reached
maturity (D = 1) in the absence of any effective measures to combat it. The
corresponding profile is assumed to be Dorrington and others' forecast for 2010.
The next step is to calculate the corresponding state probabilities .irt(st), which
requires an assumption about the incidence of the disease among couples. The
probability of transmission within a union appears to be of the order of
10 percent a year under the conditions prevailing in East Africa (Marseille,
Hofmann, and Kahn 2002), which, when cumulated over the median course of
the disease from infection to death of a decade, implies that the probability of the
event that both partners become infected, conditional on one of them getting
infected outside the relationship, is about 0.65. Given the uncertainties involved,
a less concentrated pattern of mortality within families has been assumed,
namely that the incidence is independently and identically distributed. The
resulting state probabilities are set out in table 2, where their values correspond
not to the actual years shown but rather to the steady states associated with each
disease environment (D = 0,l).
The appalling dimension-social, economic, and psychological--of the epi-
demic in its mature phase are plain. In the absence of AIDS, 85 percent of all
children would grow up enjoying the care, company, and support of both natural
parents, and fewer than 1 percent would suffer the misfortune of becoming full
orphans (table 2).If the epidemic is left to run unchecked, it will leave almost 20
percent of the generation born from 2010 onward full orphans, about 50 percent
will lose one parent in childhood, and a mere 30 percent or so will reach adult-
hood without experiencing the death of one or both parents. The epidemic
will also reverse the usual pattern of excess mortality among fathers-from
Bell, Devarajan, and Gersbach 67
about twice as high as among mothers to a third to a half lower. Given the
mother's special role in securing the young child's healthy development, it can
be argued that this reversal imparts additional force to the shock.
The final step is to undertake some sensitivity analysis. Since the decisive
factor in the system's dynamics is how At lies in relation to the critical, steady-
state values A*(s,K n,z), an appropriate way of investigating the robustness of
,
the calibration procedure is to examine the sensitivity of A*(.)to variations in
the parameter values estimated or derived above. The values of z, a,and are
estimated jointly, so one cannot be varied without modifying the others. Two
types of sensitivityanalysis can be performed. First, the three parameters can be
can varied within this straitjacket. Second, taking a! and as given, z (aswell
as K and n) can be varied in such a way that the whole configuration is actually
more optimistic than the one that emerged from the calibration (forexample, by
setting z in excess of 0.818). The second approach is chosen because it evaluates
the robustness of the findings over a much wider domain and allows the
parameters to take new values after 1990 (as already indicated for r; and n).
Table 3 sets out the values of A*(.)for a variety of plausible parameter values.
In keeping with the above discussion of fertility and mortality, the choices are
n = 3 and n = 4, with K = 0.860 and r; =0.338, which correspond to D = 0
and D = 1, respectively. The intermediate value r; = 0.6 represents a less dra-
matic, or waning, epidemic. In addition to the calibrated value z = 0.818,
somewhat more optimistic values can be considered, namely 0.9 and 1.0, as
well as the possibility that the future value of z may be reduced by the higher
dependency ratio that will attend higher premature adult mortality (say,
z = 0.7). Beginning with the calibrated values n =4, r; = 0.860, and
z = 0.818, equation (3) yields A*(s,.) = 2.06, 2.10, and 4.33 for s = 0, 1,
and 2, respectively, the first two of which lie comfortably below A60. Since the
fraction of one-parent households under nuclear family arrangements was a
I
modest 14 percent (see table 2), it can be assumed that the implicit burden of
supporting them and full orphans was both tolerable and actually taken up. It
follows that regardless of the family arrangements actually in force, the South
African economy had already been launched on a path toward steady-state
growth before the epidemic broke out in the early 1990s.
The reduction in fertility from n = 4 to n = 3 after 1990 has only a very slight
effect on A* under both family arrangements. The fall in n implies a smaller
weight on the term for altruism toward children, but this is just outweighed by
the correspondingly smaller claims that fewer children make on the family's
resources-whether they are raised under pooling or within a nuclear family.
Indeed, this effect is small in all the parameter constellations in table 3, which
leads to the conclusion that plausible changes in fertility do not play an impor-
tant role in determining the qualitative nature of the system's dynamics.
The other striking feature of table 3, by contrast, is the sensitivityof A*(s,.) to
r;.In all variations for r; = 0.338 (that is, D = I),A*(s,.) > =3.14, which
TABLE 3. Critical Values of A* (s, K, z, n)
points to a progressive economic collapse in the face of an undiminished con-
tinuation of the epidemic and in the absence of any countervailing intervention.
If K = 0.6 and n = 3, this fate is avoided under both family arrangements
(assuming, as above, that needy families will be supported) when z takes the
value 0.9 or higher. When z takes the calibrated value 0.818, however, the
pooling arrangement only barely escapes the trap, whereas the two-parent
nuclear family (s= 1) barely slips into it. Summing up, these results suggest
that even allowing for some uncertainty about the calibrated values of z and A60
and about the estimated value of K in the steady state corresponding to D =1,
as well as the behavior of fertility, the current course of the epidemic poses a
very real threat to the long-term growth of the South African economy.
Simulations
Three simulations of the course of the economy for the period after 1990 form
the set of benchmarks.
BENCHMARK 1: POOLING, NO AIDS. The corresponding trajectory of the variable
A,, about which all else revolves, is plotted in figure 2. As noted above, the key
feature of this story is that steady-state growth is ultimately attained. Starting
fromthe modest level of 0.5 in 1960, education becomesvirtuallyfull-timein the
generation born from 2020 onward, by which point, income per head is two-
thirds higher than in 1960, with another increase of 80 percent in the next
generation. The burden of child-dependencyis limited throughout: 0.65 adopted
children per couple in addition to the four of their own before 1990 and 0.49 in
addition to the three of their own thereafter. This is the relativelyhappy counter-
factual into which AIDS intrudes at t = 0 (1990).
Bell, Devarajan, and Gersbach 69
FIGURE 2. Comparison of Benchmarks (z= 0.818)
14
Benchmark 1
--t Benchmark2 (kappa=0.338)
12
r Benchmark2(kappa-0.6)
A Benchmark 3 (kappa=0.338)
The results are summarized in table 4, which provides a compact summary of
all three benchmarks that relate to the values of the parameters calibrated above.
From 1990onward, a representative family under pooling comprises two surviv-
ing adults and 3.49 children in the absence of AIDS and two surviving adults and
8.87 children in its presence.
BENCHMARK 2: POOLING, AIDS, AND NO INTERVENTION. If, following the full onset
of the AIDS epidemic, premature adult mortality remains at the level that
TABLE 4. Three Growth Paths for the South African Economy
NO AIDS 4 2 0 = 690 6750 = 690
Year 1 e ~ ( 0 ) 1 e Y(O! i e ~ ( 0 )
Note: All results are based on 30920.
a~i20 in 1990, householdsformedexpectationsabout adult mortality in 2020, when their
= 690:
children will have reached adulthood, that reflected the actual course of the epidemic over the
period 1990-2020, as set out in table 3; nfS0= 690is analogously defined when such expectations
are revised starting only in 2020.
yields the steady-state probabilities in table 2, the consequences of doing
nothing will be nothing short of disastrous, as seen in figure 2. Within a few
generations, the epidemic sets in train a complete collapse of both the economy
and, almost surely, the social institution of pooling. The extremely high level of
premature mortality among adults leaves the community relatively impover-
ished from the start and with an intolerable burden of dependency: each
surviving couple has to care for almost two adopted children for each one of
their own. Education is correspondingly neglected, with unrelieved child labor
(e =0) for the generation born starting in 2020. The descent into backward-
ness (1= 1)is complete by 2050, when family income is a little less than two-
thirds its level in 1960, and there are almost twice as many children for each
couple to care for. The results are summarized in table 4.
It might be argued that both variants with AIDS in table 4 constitute unduly
pessimistic estimates of the conditions prevailing in 1990-2020 and beyond in
terms of the level of mortality and the growth of long-term productivity. The
sensitivity analysis in section I1 covers this possibility, but those findings are
expanded on here.
Suppose, for example, that from 1990 onward, K.were to fall, not to 0.338 as
above, but less precipitously, to 0.6, say. Since /2*(0,K. = 0.6)< (table 3), no
collapse follows; but the system teters on the brink, with virtual stagnation
thereafter (figure 2). Turning to the growth of productivity over the long run,
prudent economic management and social integration after 1990 ought to yield
an improvement over z = 0.818. Suppose, then, that z = 1, which corresponds
to a doubling of A every generation (or 2.31 percent a year) under full-time
schooling. If K. continues at 0.338, however, the collapse that ensues is scarcely
less dramatic than that when z = 0.818 (figure 2).
BENCHMARK 3: POOLING, AIDS, AND DELAYED EXPECTATIONS.The second variant in
table 4 reflects the possibility that households will take some time to revise their
expectations. Suppose this revision does not occur until the very start of the next
generation, when the childhood experience of parental death will be vivid in the
minds of the next cohort of young adults: their firm expectations are = 690.
Suppose, further, that these expectations are realized and that this scale of
mortality persists into the future. The happy-but false-expectations about
future mortality that are formed in 1990, coupled with what is assumed to be
the generous altruism of full pooling, induce adults to invest heavily in the
children's education, despite the sharp reductions in available resources caused
by the outbreak of the epidemic. Yet, although the adults in the generation
starting out in 2020 are every bit as well endowed with human capital as they
would have been in the absence of the epidemic, their expectations concerning
their children's future are so bleak as to induce them to roll back investment in
schooling to levels not seen since the mid-20th century. The result is to send the
entire system into a progressive decline.
Bell, Deuarajan, and Gersbach 71
As reported in table 4, income per capita in benchmark 3 peaks in the period
starting in 2020, and two generations later, the fresh cohort of adults will be
scarcely more productive than their forebears in 1960. Only a revival of opti-
mism about the future and the resumption of low levels of premature adult
mortality to confirm it will stave off a complete collapse.5Note that a collapse is
possible even when the mortality shock affects only one 30-year generation,
depending upon how and when expectations are formed.
All policies are assumed to be financed by lump-sum taxes. Furthermore, the
government chooses the level of public expenditure not to optimize a classi-
cally specified intertemporal welfare function over an infinite horizon-a
problem that is almost impossible to solve in the framework-but to restore
steady-state growth and then maintain it. The policy program takes the form
of a sequence of taxes and expenditures that achieves this objective, if it is at
all feasible.
Health Policy
Health policy takes the form of spending on measures to combat the disease. For
some diseases, treatment may result in a completecure. There is no such prospect
for the victims of AIDS; but the treatment of opportunistic infections in the later
stages and the use of antiretroviral therapies can prolong life and maintain
productivity. In the present overlapping generations setting, therefore, treatment
may be thought of as reducing premature adult mortality in the probabilistic
sense.
It remains to establish the relationship between the state probabilities and
spending on combating the disease. This is accomplished by choosing a func-
tional form for the relationship between the probability of premature death
among adults, q, and the level of expenditures on combating the disease, 77,
and then making the simplifying assumption that the incidence of the disease is
independently and identically distributed. For simplicity, and erring on the side
of optimism, it is also assumed that such aggregate expenditures produce a pure
public good, so that
where q(7;D = 1) is to be interpreted as the efficiency frontier of the set of
all measures that can be undertaken to reduce q in the presence of the
d'~sease.
Very little is known about the exact shape of the function q(.), but
q(0;D = 1) should yield the estimates in table 2. A second, plausible, condition
5. The fact that false expectations can be helpful in overcoming shocks raises delicate questions about
the value of transparency in public policy in this context. They are avoided here.
is that arbitrarily large spending on combating the epidemic should lead to the
restoration of the status quo ante, that is, q(m;D = 1)= q(D = 0). For reasons
that will become clear shortly, it is desirable to choose a functional form that
not only possesses an asymptote but also allows sufficient curvature over some
relevant interval of v, so that the natural choice falls on the logistic:
Hence,
and
Thefull estimationof thefunction q(.) is described in appendix A2. The procedure
yields the values of the parameters a, b, c, and d for men, women, and both
combined, which are set out in table A2-1 for two values of the cost of saving a
disability-adjustedlife year. The associated functions q(v;D = 1) are convexto the
origin and have relatively strong curvature over the interval 77 E (300,700) (figure
A2-1).The said values depend on the annual cost (K) of a course of generic drugs.
Marseille, Hoffmann,and Kahn (2002)set K at $395. Early in 2006, however, the
annual cost of a course of generic drugs was about $200, so some might regard the
first estimate as too conservative in terms of the cost-effectivenessof treatment, as
opposed to prevention-ven though it bears stating that neither estimate makes
any allowance for the other components of highly active antiretroviraltherapy and
the threat that drug-resistant strains will proliferate when the full regime is not
rigorously followed. The subsections that follow begin with the results based on
the calibrated values of the parameters and the health-cost factor K = 395. The
robustness of these findings to changes in all these parametersare then examined.
Policy Option 1: Spending on Health Under Pooling
The results of spending on health under pooling are qualitatively striking(table5).
The optimal level of spending on combating the epidemic immediately upon
outbreak in 1990 (t = 0) is $963, which is about 4.5 percent of GDP, rising to
$1,029, or 3.6 percent of GDP, in 2020, when productivity is 30 percent higher.
Fiscally speaking, this is a tall order and a very substantial long-run burden,
especiallyin view of the fact that the additional taxes are assumed to be raised in
lump-sum form6 If this program is politically feasible, it will eventually yield
6. Under the distortionary tax systems that rule in practice, the marginal cost of a unit of public
revenue can range between 1.3 and 1.7 or higher still.
Bell, Devarajan, and Gersbach 73
TABLE 5. Policy Option 1: Spending on Combating the Disease under Pooling
- - -
Year (1960) l(2.62) e (0.50) 1 (0) K (0.860) n / (4.65)
~ y(0)(19,503)
- - -
Note: 1 is the per household level of spending on combating the disease; y(0)is the level of
income accruing to each pair of surviving adults and the children in their care.
steady-state growth, with full and universal education attained in 2050. With
(optimal) spending at this level, premature mortality among adults would be
scarcely higher than in the complete absence of the disease.
A comparison with benchmark 1 reveals that the costs of dealing with AIDS in
terms of lost output are modest at first but become quite large by 2080, when
productivity is about 88 percent of its benchmark level, even with the optimal
package of interventions under the favorable conditions of the case considered
here (table4). The long-run rate of growth is unaffected by AIDS under this policy
program; for once full-time schooling is reached, the growth rate depends only
on z(0,K), which is assumed to be constant at z(0,0.86)= 0.818. Taking a
somewhat broader view, therefore, the outcome is encouraging, in that the
general character of benchmark 1 is still attainable (figure 3), including a
relatively low level of premature adult mortality. Thus, the maintained assump-
tion that pooling will survive the shock is arguably validated.
Given this rather encouraging qualitative finding, it is still natural to ask
whether lower costs of generic drugs will yield significant quantitative gains.
When K = 200, the optimal values of in 1990 and 2020 and thereafter are
substantially lower at $714 and $755, respectively, but the corresponding
values of K still rise, to 0.854 and 0.855, respectively. With a much lower
fiscal burden and a slight improvement in premature adult mortality,
A increases a little more rapidly than when K = 395, so that its level in 2080
is almost 4 percent higher.
Policy Option 2: Nuclear Families, Lump-Sum Subsidies
The results under policy option 1 are predicated on the assumption that the
government acts at once to nip the epidemicin the bud. In fact, the epidemichad
assumed alarming proportions by 2000, with many children already left as
74 T H E W O R L D B A N K E C O N O M I C REVIEW, VOL. 20, N O . I
FI GURE 3. Policy Options 1, 2, and 3 (K = 395)
orphans and even more destined to become orphans, thus calling into question
the whole system of pooling. If this social institution does break down, leaving
tightly defined nuclear families to emerge instead, then the government will face
the challenging task not only of averting a collapse, but also of preserving
equality within each generation. To make both possible, additional assumptions
are needed about the formation of human capital when children are left as half
or full orphans. Under the assumption that z(1) = z(2)/2, that is, single parents
can do just as well as couples in raising their children if they have the income?it
is possible to preserve equality of educational outcomes among all children with
at least one living parent by subsidizing one-parent families so as to induce them
to choose the same level of education that two-parent families choose. By
hypothesis, no family takes in full orphans, so that these children must be
cared for in orphanages. It is assumed that these institutions, when properly
staffed and run, substitute perfectly for parents, at least where the formation of
human capital is concerned. The operating rule is that each full orphan also
enjoys the same level of consumption as a child in a single-parent household.
When the family structure is nuclear, a good policy program to overcome the
shock caused by AIDS must ensure a substantial tax base, not only in the present
but also in the next generation. The instruments available for this purpose are
taxes on two-parent households, spending on combating the disease, the size of
the subsidy to single-parent households, and the proportions of half and full
orphans to be supported. They are chosen subject to the above restrictions
Bell, Devarajan, anal Gersbach 75
designed to preserve equality, if at all possible, and to the government's budget
constraint.
Given the complexity of using full-scale forward induction, a somewhat
simpler approach is chosen. The aim is to maximize the expected size of the
tax base in the next period, where all parties hold the firm expectation that there
will be a continuation of the level of premature adult mortality (andhence of 17)
prevailing in the present. That is, stationary expectations are assumed, which
permit the maximization problem to be written so that it effectivelycontains no
variables or parameters pertaining to the future. In particular, families' deci-
sions about education depend on but under stationary expectations,
= K ~ .
The (bounded)rationality of these expectations is secured by impos-
ing the condition that 17 does not fall from one period to the next, for this will
rule out policy programs under which the value of investments in education will
be reduced ex post by failures to take adequate measures against the disease in
the next period. It should be emphasized that if it is possible to stave off a
collapse of the economy through a policy program derived on the basis of
stationary expectations so formulated, then it certainly will be possible to do
even better by using the full apparatus of forward induction. Since all adults
possess at least one unit of human capital, the tax base is defined, for present
purposes, as the excess of the aggregate level of human capital over the aggre-
gate level when all adults have but one unit.
The optimum sequence7yieldsa continuation of growth with complete equal-
ity-all orphans receive the support needed to bring them up to par with the
children of two-parent households in each and every period (table 6). Growth is
distinctly sluggish, however, which points to a collapse that would be somewhat
narrowly averted. The (uniform)years of schooling rise noticeably more slowly
across succeeding generations than under pooling, with full-time schooling
achieved only in 2080, when the level of productivity is only slightly more
than double its value in 1990. Spending on combating the disease is also higher
in absolute terms throughout and, combined with the transfers required to
support needy children, g(2), this yields a much heavier fiscal burden than
under pooling. Two-parent households pay a little over 20 percent of their
income in the form of a lump-sum tax (7) to finance this program in 1990
and receive very little relief until rapid growth begins from 2080 onward, when
one-parent families need less support.
The differences between policy options 1 and 2 call for some explanation.
Under pooling, which ensures equality, the objective is to maximize the
(uniform) level of productive efficiency (A)in the next generation, whereas
with nuclear families, it is the size of the future tax base that matters when the
government has to undertake the task of replacing the institution of pooling
with subsidies and orphanages. In the latter arrangement, it may be worth-
while to trade off educational attainment to secure more surviving adults at
7. The optimization problem is set out in full in Bell, Devarajan, and Gersbach (2003).
TABLE 6. Policy Option 2: Nuclear Families, Lump-Sum Subsidies
Year
Note: g(2)is the income transfer to each one-parent family that receives such support; z is the
level of the special lump-sum tax on each two-parent family; 6(s), s = 2,4 is the fraction of all
children in family state s receiving public support; y(1) is the level of gross income accruing to a
two-parent family.
the later date. That is exactly what has happened here: the absolute level of rl is
14 percent higher than under pooling in both 1990 and 2050, despite the fact
that productivity under pooling is 57 percent higher in the latter period. The
other contributing factor arises from the fact that raising children in.orpha-
nages draws some adults out of the production of the aggregate private good-
a cost that does not arise (by assumption) under pooling. The upshot is that
families have less disposable income than under pooling, so that their children
receive fewer years of schooling and growth is much slower. As under pooling,
the long-run rate of growth is unaffected by AIDS in this fairly good sequence;
but the traverse to steady-state growth is a painfully long one.
How much less painful would this trek be when K = 200? As under
pooling, the optimal levels of rl are just over 25 percent lower than when
K = 395, and K edges up further, almost to what its level would be in the
absence of the epidemic. The absolute tax burden on two-parent families is
also somewhat lighter: 8.6 percent lower in 1990, 7.5 percent in 2020, 2.6
percent in 2050, and almost 36 percent lower in 2080, when single-parent
families need much less income support to be induced to choose full-time
schooling. The effects on the accumulation of human capital are small at
first, but by 2080, 2 is 14 percent higher than when K = 395. Since full
equality in terms of human capital within each generation emerges as part
of the optimal program, this faster pace requires that one-parent families
need more generous support up to 2080, and g(2) is correspondingly more
generous-5.5 percent higher in 1990,7.5 percent in 2020, and 11 percent in
2050.
Bell, Devarajan, and Gersbach 77
Policy Option 3: Nuclear Families, School Attendance Subsidies
The results for this option are qualitativelysimilar to those under policy option 2,
but growth is considerably more rapid (table 7; figures 3 and 4). Given the
efficiency of school attendance subsidies relative to lump-sum transfers (and
hence the lower taxes on two-parent households), one would expect a swifter
attainment of full-time schooling- in ths variant, and this is indeed the case here.
The precise reasoning runs as follows. Choose the optimal levels of taxes on two-
parent households and spending on health under policy option 2. This program
will yield thesame demographicstructure, the samelevel of education among such
families, and the same total tax revenues. The outlays under policy option 3
needed to induce the same level of education among the children of one-parent
households, however, will be smaller than under policy option 2. These children
will also have a lower level of consumption, a standard to which full orphans are
tethered. It follows that there will be an excess of total revenue over expenditures.
Let rl be held constant, so as to keep the demographic structure unchanged, and
let the taxes on two-parent households be reduced slightly, which will induce a
small rise in e(1).By continuity, there will still be enough funds to finance the
additional subsidies to half and full orphans that will be needed to preserve
equality in education, and hence in human capital in the next generation of
adults. It follows that policy option 3 strictly dominates option 2 in all periods
from t = 0 onward.
Full education is reached in 2050, as is the case under pooling, although
productivity is 12 percent lower, because of the accumulated effects of
lower attainments in the two preceding generations. Spending on measures to
combat the disease is a little higher than under pooling at first but a little less
from 2020 onward. It is about 13 percent lower than its co~unterpartunder
policy option 2 throughout, so that more premature deaths are implicitly
accepted, although the differences in K, are small. A measure of the comparative
efficiency of conditional educational subsidies is that satisfa.ctory growth is
TABLE 7. Policy Option 3: Nuclear Families, School Attendance Subsidies
`Year 1 4 1 ) 4 2 ) v g(2) .r 4 2 ) 4 4 ) K ~ ( 1 )
78 T H E W O R L D BANK E C O N O M I C REVIEW, VOL. 20, N O . I
FI GURE 4. Policy Options 1,2, and 3 (K = 200)
achieved with amounts paid to one-parent households that are barely a tenth of
the lump-sum transfers made under policy option 2.
The tax burden on two-parent households is correspondingly lighter: the
absolute payment per household is a little less than one-half of that under policy
option 2 in 1990, rising to 56 percent in 2080. The difference in productivitiesis
very large in 2080: namely 241) to one, or 1.636, which implies a much lower
relative tax burden. The latter falls from about 8.6 percent of income in 1990 to
3.6 percent in 2080 under policy option 3, and from 19.3 percent to 10.4
percent under policy option 2. Observe that although the payment of school
attendance subsidies ends from 2080 onward, T is higher than in 2050. The
reason is that the raising and caring for full orphans require the time and effort
of adults specifically employed for this purpose, the costs of which rise with 1.
An Optimistic Variant
This analysis of alternative policies concludes with brighter assumptions about
the long-term rate of growth of productivity and the costsof antiretroviral drugs.
The results for z = 1 and K = 200, perhaps the most plausible of optimistic
constellations, are reported in tables 8-10. Table 8 sets out the first benchmark
(D = 0). In the absence of the epidemic, the increase in z from 0.818 to 1 makes
for large gains indeed over three generations. Full schoolingis achieved in 2020,
and 2 in 2080 is almost 75 percent higher than in the reference case. Turning to
policy interventions in the face of the epidemic, the optimal value of is not
Bell, Devaraian, and Gersbach 79
TABLE 8. An Optimistic Variant (z =1, K = 200): The First Benchmark (D= 0)
TABLE 9. Policy Option 1: Spending on Combating the Disease under Pooling
(z = 1, K = 200)
-
Year i e II K n / ~ ~ ( 0 )
TABLE 10. Policy Options 2 and 3: Nuclear Families, Lump-Sum and School
Attendance Subsidies (z = 1, K = 200)
Year 1 4 1 ) 4 2 ) v g(2) 7 4 2 ) 44.) K y(1)
Policy option 2
1990
2020
2050
2080
Policy option 3
1990
'2020
2050
2080
affected under pooling, and 1 in 2080 is a mere 3 percent smaller than the
corresponding benchmark value. With nuclear families, the pace is also dis-
tinctly quicker, even under policy option 2. A more efficient educational tech-
nology will do much to ease the task of maintaining growth and welfare in the
face of the epidemic, but it will not necessarilystave off a collapse in the absence
of any other intervention (table 3).
80 T H E W O R L D BANK E C O N O M I C REVIEW, VOL. 20, N O . I
The AIDS epidemic will peak far in advance of the economic damage it will
ultimately cause. In southern Africa, where prevalence rates among people
aged 15-49 years are already 20 percent and higher, the worst is still to come.
The scale of that damage, in terms of accumulated losses in GDP per capita, will
also be large even if the measures designed to combat the disease and to ensure
the education of orphans are well chosen, and the fiscal means employed to
finance them are highly efficient. Without such measures, and given a continua-
tion of high levels of mortality, economic collapse is a very real danger.
The main reason for these gloomy findings lies in the peculiarly insidious and
selectivecharacter of the disease. By killing mostly young adults, AIDSdoes more
than destroy the human capital embodied in them; it deprives their children of
the very things they need to become economically productive adults-their
parents' loving care, knowledge, and capacity to finance education. This weak-
ening of the mechanism through which human capital is transmitted and accu-
mulated across generations becomes apparent only after a long lag, and it is
progressively cumulative in its effects. Therein lies the source of the difference
between the findings in this article and those of many previous studies, which
have focused either on the role of quasi-fixed factors over the medium run or on
the historical record to date.
What are the lessons for public policy?Where the prevalence rate is still low,
as in much of Asia, Eastern Europe, the Middle East, and Latin America, it is of
the utmost importance to contain the disease at once: for the economic system
as well as for individuals, an ounce of prevention is worth more than a pound
of cure.
Where the epidemic is more advanced, combating the disease and its economic
effects successfully will require a large and determined fiscal effort, the correct
design of which is a complicatedmatter. Intuitively, the question is: What combi-
nation of measures should be adopted to promote the formation of human capital
and good health when the threat of a collapse looms? These measures are partly
complementary.Maintaininggood health means that the human capital embodied
in individuals during childhood and training will survive and pay off into old age,
not only for them but also for their children. When public funds are very scarce,
however, some tradeoffs will be unavoidable, requiring the concentration of
resources on some programs or groups at the expense of others. The hope here
is that the knowledge about what works in the fields of child rearing, education,
the care of orphans, health, and so forth can reveal how to formulate combined
programs of interventions that will ward off the threat of an economic collapse.
The true social rate of return to such programs can be extremely high, whereas
that derived from calculations based on standard (local)cost-benefit analysis may
be quite modest. Fiscal policy in general, and policy in the social sectors in
particular, must be formulated with a clear eye on its contribution to solving the
Bell, Devarajan, and Gersbach 81
long-run economic problem posed by AIDS. For in the event of a collapse of
productivity, little else will matter.
These points are vividly illustrated by the results for South Africa. In the
absence of the epidemic, there would have been the prospect of modest, but
accelerating growth of per capita income. An unabated con1:inuation of the
epidemic could bring about a progressive collapse. With the right interventions,
this fate can be averted, although the costs are high, even under favorable social
arrangements for thecare of orphans. If those arrangements break down, growth
is likely to be rather sluggish. These conclusions must be regarded as prelimin-
ary, and various aspects of the analysis need further work and refinement. That
much conceded, the sensitivity analysis nevertheless suggests that these findings
are robust to changes in a variety of key assumptions and parameter values. And
it would be unconscionable to err on the side of optimism.
APPENDIX A l . MICROFOUNDATION OF THE MODEL
This appendix presents the microfoundation of the dynamical system. It does not
attempt to model individuals' sexual behavior, with all its gratifications and
risks. While the said probabilities are therefore exogenous, their values in the
application to South Africa are based on the epidemiological work of Dorrington
and others (2001).
The first topic is the formation of human capital. Consider a family at the
start of period t. Let 1$ be the father's endowment of human capital, A Tthe
mother's, and At(st)their total human capital when the family is revealed to be
in state st. Then,
Assuming that there is assortative mating, A{ = AT, equation (Al.l)specializesto
(A1.2) At(1) =2At,At(2) =At(3) = At, At(4) =0,
where the superscripts f and m may now be dropped without introducing
ambiguity.
Human capital is assumed to be formed by a process of child-rearing com-
bined with formal education. In the course of rearing their children, parents give
them a certain capacity to build human capital for adulthood, a capacity that is
itself increasing in the parents' own human capital. This gift will be of little use,
however, unless it is complemented by at least some formal education, in the
course of which the basic skills of reading, writing, and calculating can be
learned. Let the proportion of childhood devoted to education be denoted by
et E [O, I], the residual being allocated to work, and for simplicity, let all the
children in a family be treated in the same way.*The human capital attained by
each of the children on reaching adulthood is given by
Beginning with the upper branch of equation (A1.3),the term zt(st)represents
the strength with which capacity is transmitted across generations. For simpli-
city, the father's and mother's contributions are assumed to be perfect substi-
tutes: 4 2 ) = ~ ( 3 )It. is also assumed that where transmitting this capacity is
concerned, two parents can rear a child at least as well as one, but, in view of
perfect substitutability, no better than twice as well as one. Hence, recalling
equation (A1.2),
Thus, the upper branch of equation (A1.3)can be rewritten as
with both types of single-parent families being identical in this respect. The
function f (.) represents educational technology-translating time spent on edu-
cation into learning. It is assumed to be strictly increasing and differentiable,
with f (0)= 0. Observe that equation (A1.3)and f (0)= 0 imply that children
who do not attend school at all attain, as adults, only some basic level of human
capital, which has been normalized to unity.
According to the lower branch of equation (A1.3), there is a miserable out-
come for full orphans who do not enjoy the good fortune to be adopted or placed
in (good) institutional care. Deprived of love and care, and left to their own
devices, they go through childhood uneducated, to attain human capital < (5 1)
in adulthood.
The next step is to relate human capital to current output, which takes the
form of an aggregate consumption good. Output is assumed to be proportional
to inputs of labor measured in efficiency units. A natural normalization is that an
adult who possesses human capital in the amount It is endowed with At effi-
ciency units of labor, which he or she supplies completely inelastically. A child's
efficiencywill be somewhat lower than the parents', all other things being equal,
on the grounds of age alone. To reflect these considerations, let a child supply
(1- et)yefficiencyunitsof labor when the childworks1- et units of time. It is
8. This analysis skips the fact that girls often receive less education than boys. The ensuing inequality
in human capital introduces analytical and empirical difficultieswhose importance, for the purpose of this
article, does not seem to warrant specific treatment.
Bell, Devarajan, and Gersbach 83
plausible to assume that y E (0, t),that is, a full-time working child is at most as
productive as an adult who happened to be an uneducated orphan. A family
with n, children therefore has a total income in state st(st= 1,2,3)of
where the scalar a(>0) denotes the productivity of human capital, measured in
units of output per efficiency unit of labor input.
Household Behavior
All allocative decisions are assumed to lie in the parents' hands, as long as they
are alive. Any bequests at death are ruled out, so that the whole of current
income, as given by equation (A1.6),is consumed. Within the family, let the
husband and wife enjoy equality as partners, and let each child obtain a
fraction ,/3 E (0,l) of an adult's consumption if at least one adult survives.
Full orphans (st = 4) do not attend school and consume what they produce as
child laborers.
Without any taxes or subsidies, the household's budget constraint may there-
fore be written as
(A1.7) [(3- st) + ntP]c,+ antyet 5 a[(3 +
- st)& n,y], st = 1 , 2
where c, is the level of each adult's consumption. The expression on the left
hand side represents the costs of consumption and the opportunity costs of the
children's schooling. The expression on the right hand side is the family's so-
called full income9 in state st= 1,2,3. Observe that single-parent households
not only have lower levels of full income than their otherwise identical two-
parent counterparts, but they also face a higher relative price of education,
defined as anty/[(3- st) +nt,/3].
Couples have children while they are young until some exogenously fixed
number have survived infancy, a target that may vary from period to period.
With nt thus fixed,'' the adults wait until the state of thefamily becomes known,
and the survivor then chooses some feasible pair (c,, et) subject to condition
(A1.7).
Parents are assumed to have preferencesover their own current consumption
and the human capital attained by their children in adulthood, taking into
account the fact that investment in a child's education will be wholly wasted if
that child dies prematurely in adulthood. Let mothers and fathers have identical
preferences, and for two-parent households, let there be no joint aspect to the
consumption of the pair (c,,e,): each surviving adult derives (expected) utility
9. A household's full income is the scalar product of its endowment vector and the vector of market
prices. Here, output is taken as the nurniraire.
10. Although there is much evidence in favor of at least some replacement fertility, this is evidently a
strong assumption. In the numerical application, however, variations in n,turn out to have only weak
effects on the system's dynamics (seetable 4).
from the pair so chosen, and these utilities are then added up within the family.
In effect, whereas ct is a private good, the human capital of the children in
adulthood is a public good within the marriage.
Since all the children attain At+l, the only form of uncertainty is that
surrounding the number who will not die prematurely as adults, which is
denoted by the random variable at+l. Let preferences be separable, with
representation
where the contribution V(A,+~)counts only when death does not come
early, Et is the expectation operator, and Etat+1is the expected number of
children surviving into old age. The subutility functions u(.) and v(-)are
assumed to be increasing, continuous, concave, and twice-differentiable.
Denoting by K:, the parents' subjective probability that a child will survive
to old age and recalling assumption 1 and that all children are treated
identically yield
where it+l is given by equation (A1.3).A reduction in ~ f +therefore effectively
~
entails a weaker taste for children's education. It will be convenient in what
follows to rewrite equation (A1.8)as
since both types of single-parent families are identical. Hence, it suffices to
examine the states st = 1,2. A family in state st (= 1,2) in period t solves the
following problem:
( A l . l l ) max EUt(st) s.t. (A1.7),ct > 0,et E [O,l].
[s(st),et(st)l
Let [c:(st),e: (st)] solve problem ( 11 1 , whose parameters are
(a,p, y,K:+~, At, nt).Using the envelope theorem yields
Since current consumption is maximized by choosing et = 0, it follows that the
parents' altruism toward their children must be sufficiently strong if they are to
choose et > 0.
Bell, Devarajan, and Gersbach 85
If both goods are noninferior, it follows at once that
Dynamics
There are no insurance arrangements in the above account, so that premature
adult mortality in period t will affect not only the level but also the distribution
of human capital in ~eriodt +1. As noted above, full orphans will suffer low
~roductivityin adulthood, as expressed in the lower branch of equation (A1.3).
Such mortality also affects the distribution of families across states 1,2, and 3 in
period t and will thus affect the level and distribution of human capital in period
t+ 1 if e:(st) varies across states and with the severity of premature adult
mortality, as it normally will when At is not too large. These repercussions
will then make themselves felt in future periods, even if premature adult mor-
tality vanishes after period t.
To state all this formally, recall that the family chooses e;(st;.) in light of its
resources and expectations so as to solve problem (Al.ll). Hence, equation
(A1.3) may be written so as to make these influences explicit:
Equation (A1.12)describes a random dynamical system-random in the sense
that each child in any given family state st can wind up in any of the states
st+l E {1,2,3,4) after reaching adulthood and forming a family in period t 1. +
With four parameters to be estimated, two additional independent conditions
beyond those in the text are required. One way of proceeding is to pose the
question: What is the marginal effect of efficient spending on q in high-and low-
prevalence environments? That is to say, estimates are needed of the derivatives
of q(q;D =1)at q = 0 and some value of q that corresponds to heavy spending,
when the scope for exploiting cheap interventions has been exhausted. To
obtain such estimates, we used the estimated costs of preventing a case of AIDS
or saving a disability-adjusted life year by various methods, as reported by
Marseille, Hofmann, and Kahn (2002).
When the prevalence rate is high, the authors argue, the most cost-efficient
form of intervention is to target prostitutes for the specific purpose of controlling
sexually transmitted diseases and promoting the use of condoms. The associated
cost per AIDS case averted in Kenya is given as $8-12. It seems reasonable to infer
that this cost recurs annually. Other preventive measures are less cost-effective
by a factor of up to 10 or more. Marseille, Hofmann, and Kahn (2002)put the
average cost per disability-adjusted life year of a diverse bundle of such measures
at $12.50. For these measures, the assumption that 77 produces a pure public
good is not far off the mark. Now, a reduction in q of 0.01 over a span of 30
years yields 0.3 disability-adjusted life years. Allowing for the fact that there is
substitution among diseases, that is, if one does not succumb to AIDS, there
is always the threat of something else, the expenditure of another $12.50
when 71 is small will yield a net reduction in q(D =1) of about
(0.01) . (110.3). [I- q(D = O)]= 0.028. Recalling that 77 is defined with refer-
ence to a population of adults whose measure has been normalized to unity and
rounding up to $15, we have
Following the purposive and determined implementation of the full battery of
preventive measures, the remaining intervention is to treat the infected. There is
now neither a cure nor the prospect of one for perhaps decades to come.
Opportunistic infections can, of course, be treated in the later stages of the
disease, and the onset of full-blown AIDS can be delayed for some years through
the controlled use of antiretroviral therapies. Such measures will do little to
reduce q as strictly defined, but by keeping infected individuals healthier and
extending life a bit, they will raise lifetime income and improve the parental care
enjoyed by children in affected families.
In the context of the model, therefore, it seems perfectly defensible to
interpret these gains as equivalent to a reduction in q. Marseille, Hofmann,
and Kahn (2002)put the cost of saving a disability-adjusted life year by such
means at $395, assuming that the drugs take the form of low-cost generics
and explicitly neglecting the costs of the technical and human infrastructure
needed to support an effective, so-called highly active antiretroviral therapy
regimen of this kind. It is assumed here that the highly active antiretroviral
therapy regimen is the efficient, marginal form of intervention when a low
prevalence rate has resulted from a determined, extensive, and continuing
effort at prevention. To complete the specification of this case, the level of
aggregate spending at which highly active antiretroviral therapy becomes the
best choice at the margin must be determined. Note that in the absence of
diminishing returns to preventive measures, it would be possible to attain the
status quo ante (D = 0) by spending
In fact, diminishing returns will set in as the prevalence rate falls. Where
preventing mother-to-child ,transmission is concerned, for example, a drop in
the prevalence rate from 30 percent to 15percent will almost double the cost of
Bell, Devarajan, and Gersbach 87
saving a disability-adjustedlife year (Marseille,Hofmann, and Kahn 2002). Since
15 percent hardlycounts as a low level of prevalence,it seemsfairly safe to assume
that highly active antiretroviral therapy will not become cost-efficient until spend-
ing on preventive measures, and the treatment of opportunistic infections is at
least triple the above estimate. Thus the required fourth condition is:
The four conditions (8), (9), (A2.1), and (A2.3) may be solved to yield the
values of the parameters a,b,c, and d for men, women, and both combined, as
set out in table A2-1. Premature adult mortality is defined precisely as death
before age 50, conditional on surviving to age 20, the corresponding probability
being denoted as 30920.
By way of sensitivityanalysis, to K, the function q(.)is respecified as follows:
where K denotes the annual cost of a course of generic drugs. Equations (A2.1)
and (A2.3) are then modified to read
and
respectively. Setting K = 200 yields the parameters reported in the lower half of
Table A2-1 and the associated function in figure A2-1. The shift of q(.) repre-
sents the favorable effects of the said reduction in costs on the government's
possibilitiesof reducing premature adult mortality.
TA BLE A2-1. Parameters of the q(.)Function
a b c d
K = 395
Women 0.6613 0.0051 0.4464 1.6101
Men -0.6555 0.0034 0.1451 -1.3432
Average 0.3562 0.4396 0.8145 2.9450
K = 200
Women 1.0532 0.0169 1.8875 1.0473
Men 0.6469 0.0118 0.2524 1.7280
Average 0.1633 0.0073 0.0152 6.2653
8 8 THE W O R L D BANK E C O N O M I C REVIEW, VOL. 20, N O . I
FI GURE A2 -1. Premature Adult Mortality 30q20 as a function of 77 and K
0'8
I I I I I I I
0 200 400 600 800 1000 1200 1400
Spendingon AIDS
Arndt, C., and J. Lewis. 2000. "The Macro Implications of HIV/ATDSSouth Africa: A Preliminary
in
Assessment." South African Journal of Economics 68(5):856-87.
Barro, R., and J. W. Lee. 1996. "International Data on Schooling Years and Schooling Quality." American
Economic Review, Papers and Proceedings 86(2):218-23.
Bell, C., and H. Gersbach. 2001. "Child Labor and the Education of a Society." IZA Working Paper 338.
Institute for the Study of Labor, Bonn, Germany.
Bell, C., S. Devarajan, and H. Gersbach. 2003. "The Long-run Economic Costs of MDS: Theory and an
Application to South Africa." Policy Research Working Paper 3152, Washington, D.C.
Bell, C., S. Devarajan, and H. Gersbach. 2004. "Thinking about the Long-run Economic Costs of Hrv/
AIDS." In M. Haacker, ed., The Macroeconomics of HIV/AIDS. Washington, DC: IMP.
Bloom, D., and A. Mahal. 1997. "Does the AID Epidemic Threaten Economic Growth." Journal of
S
Econometrics 77(1):105-24.
Bloom, D. 2001. "AIDS and Economics," Mahal, A., Sevilla, J. and River Path Associates, paper
prepared for Working Group I of the WHO Commission on Macroeconomics and Health,
November 2001.
Bmhns, R. 2005. "The Long-run Effects of HW/AIDS Kenya." South-Asia Institute, University of
in
Heidelberg.
Case, A., C. H. Paxson, and J. D. Ableidinger. 2002. "Orphans in Africa." Working Paper 9213. National
Bureau of Economic Research: Cambridge, MA.
Corrigan, P., G. Glomm, and F. MCndez. 2004. "AIDS, Human Capital and Growth." Indiana University,
Bloomington, IN.
Bell, Devarajan, and Gersbach 89
Corrigan, P., G. Glomm, and F. Mindez. 2005. "AIDS Crisis and Growth." Journal of Development
Economics 77(1):107-24.
Cuddington, J. T. 1993. "Modeling the Macroeconomic Effects of AIDS, with an Application to Tanzania."
The World Bank Economic Review 7(2):173-89.
Cuddington, J. T., and J. D. Hancock. 1994. "Assessing the Impact of AIDS on the Growth Path of the
Malawian Economy." Journal of Development Economics 43(2):363-68.
Dorrington, R., D. Bourne, D. Bradshaw, R. Laubscher, and I. M. Timseus. 2001. "The Impact of HN/MDS
on Adult Mortality in South Africa." Technical Report, Medical Research Council, Tygerburg, South
Africa.
Dorrington, R., and D. Scheider, D. 2001. "ASSA2000 Urban-Rural Aids and Demographic Model."
Convention Papers 2001, Actuarial Society of South Africa, Cape Town, South Africa.
Ferreira, P. C., and S. Pessoa. 2003. "The Long-Run Economic Impact of AIDS." Fundacao Getulio
Vargas, Rio de Janeiro, Brazil.
Gertler, P. J., D. Levine., and S. Martinez. 2003. "The Presence and Presents of Parents: Do Parents
Matter for More than their Money?"University of California at Berkeley, Berkeley.
Hamoudi, A., and N. Birdsall. 2004. "HIV/AIDSand the Accumulation and Utilization of Human Capital in
Africa." In M. Haacker, ed., The Macroeconomics of HIVIAIDS, Washington, D.C.: International
Monetary Fund.
Kambou, G., S. Devarajan, and M. Over. 1992. "The Economic Impact of AIDS in an African Country;
Simulations with a CGE Model of Cameroon." Journal of African Economies 1:109-30.
Marseille, E., P. B. Hofmann, and J. G. Kahn. 2002. "HIV Prevention before HAARTin Sub-Saharan Africa."
The Lancet 359:1851-6.
McDonald, S., and J. Roberts. 2004. "~~Ds'andEconomic Growth: A Human Capital Approach." Sheffield
Economic Research Paper Series, SERP 2004008.
Over, A. M. 1992. "The Macroeconomic Impact of AIDS in Sub-Saharan Africa." AF~PNTechnical Working
Paper 3, Population, Health and Nutrition Division, Africa Technical Department, World Bank,
Washington, D.C.
UNAIDS. 2004. "2004 Report on the Global AIDS Epidemic." Geneva: United Nations.
World Bank. 2002. World Development Indicators, Washington, D.C: World Bank.
Young, A. 2005. "The Gift of the Dying: The Tragedy of AIDS and the Welfare of Future African
Generations." Quarterly Journal of Economics 120(2):423466.
Robust Multidimei~sionalSpatial Poverty
-
Comparisons in Ghana, Madagascar, and Uganda
lean-Yves Duclos, David Sahn, and Stephen D. Younger
Spatial poverty comparisons are investigated in three African countries using multi-
dimensional indicators of well-being. The work is analogous to the univariate
stochastic dominance literature in that it seeks poverty orderings that are robust
to the choice of multidimensional poverty lines and indices. In addition, the study
seeks to ensure that the comparisons are robust to aggregation procedures for
multiple welfare variables. In contrast to earlier work, the methodology applies
equally well to what can be defined as "union," "intersection," and "intermediate"
approaches to dealing with multidimensional indicators of well-being. Furthermore,
unlike much of the stochastic dominance literature, this work computes the sam-
pling distributions of the poverty estimators to perform statistical tests of the
difference in poverty measures. The methods are applied to two measures of
well-being, the log of household expenditures per capita and children's height-for-
age z scores, using data from the 1988 Ghana Living Standards Study survey, the
1993 National Household Survey in Madagascar, and the 1999 National Household
Survey in Uganda. Bivariate poverty comparisons are at odds with univariate compar-
isons in several interesting ways. Most important, it cannot always be concluded that
poverty is lower in urban areas in one region compared with that in rural areas in
another, even though univariate comparisons based on household expenditures per
capita almost always lead to that conclusion.
It is common to assert that poverty is a multidimensional phenomenon, yet most
empirical work on poverty, including spatial poverty, uses a unidimensional
yardstick to judge a person's well-being, usually household expenditures or
income per capita or per adult equivalent. When studies use more than one
indicator (ofwell-being, poverty comparisons are either made independently for
Jean-Yves Duclos is a professor of economics and director of the Inter-University Center on Risk,
Economic Policies, and Employment (CIRPBE) at Universiti Laval; his email address is jyves@ecn.ulaval.ca.
David Sahn is a professor of economics and director of the Food and Nutrition Policy Program at Cornell
University; his email address is desl6@cornell.edu. Stephen D. Younger is an associate director of the
Food and Nutrition Policy Program at Cornell University; his email address is sdyl@cornell.edu. The
authors are grateful to three anonymous referees and the editor for comments on a previous draft. The
research for this study is supported by the Strategies and Analysis for Growth and Access project, funded
by a U.S. Agency for International Development cooperative agreement with Cornell University and
Clark-Atlanta University and by the Poverty and Economic Policy network of the International Develop-
ment Research Centre. For more information, see http://www.saga.cornell.edu and www.pep-net.org/.
THE WORLD BANK ECONOMICREVIEW, VOL 20, NO. 1, pp. 91-113
. doi:10.1093/wber/lhj005
Advance Access publication April 6, 2006
O The Author 2006. Published by Oxford UniversityPress on behalf of the International
Bankfor Reconstruction and Development ITHE WORLD BANK. All rights reserved.For permissions,
please e-mail: joumals.permissions@oxfordjoumals.org.
each indicator1 or made using an arbitrarily defined aggregation of the multiple
indicators into a single index2In either case, aggregation across multiple welfare
indicators and across the welfare statuses of individuals or households requires
specific aggregation rules that are necessarily arbitrary.3Multidimensional
erty comparisons also require the estimation of multidimensional poverty lines, a
procedure that is problematic even in a unidimensional setting.
Taking as a starting point the conviction that multidimensional poverty
comparisons are ethically and theoretically attractive, the purpose here is
to apply quite general methods for multidimensional poverty comparisons to
the particular question of spatial poverty in three African countries-Ghana,
Madagascar, and Uganda. The relevant welfare theory and accompanying
statistics are developed elsewhere (Duclos, Sahn, and Younger 2003). The
purpose here is to give an intuitive explanation of the methods and to show
that they are both tractable and useful when applied to spatial poverty in
Africa.
The poverty comparisons use the dominance approach initially developed by
Atkinson (1987)and Foster and Shorrocks (1988a, 1988b, 1988c)in a unidi-
mensional c o n t e ~ tIn~ a review of this literature, Zheng (2000) distinguishes
.
between poverty comparisons that are robust to the choice of a poverty line and
those that are robust to the choice of a poverty measure or index. Both are
attractive features of the dominance approach because they enable the analyst to
avoid relying on ethically arbitrary choices of a poverty line and a poverty
measure. The poverty comparisons used here are robust to the selection of
both a poverty line and a poverty measure. In the multidimensional context,
this includes robustness over the manner in which multiple indicators interact to
generate overall individual well-being.
Section I briefly presents the data and provides an intuitive discussion of
multidimensional poverty comparisons. In addition to the stochastic dominance
conditions that are familiar from the univariate literature, it discusses two
concepts that arise only in a multivariate context. First, it distinguishes between
intersection and union definitions of poverty.5 By the well-known focus axiom
used in poverty measurement (see, for instance, Foster 1984),these definitions
1. This would involve, say, comparing incomes across regions and then comparing mortality rates
across regions and so on.
2. The best-known example is the human development index of the United Nations Development
Programme (UNDP 1990),which uses a weighted average of life expectancy, literacy, and GDP per capita
across the population.
3. Such rules have been the focus of some of the recent literature. See, for instance, Tsui (2002)and
Bourguignon and Chakravarty (2003).Bourguignon and Chakravarty (2002)also give several interesting
examples in which poverty orderings vary with the choice of aggregation rules.
4. Atkinson and Bourguignon (1982, 1987) first used this approach in the context of multidimen-
sional social welfare. See also Crawford (1999).
5. For further recent discussion, see Bourguignon and Chakravarty (2002, 2003), Atkinson (2003),
and Tsui (2002).
Duclos, Sahn, and Yozrnger 93
identify the individual poverty statuses to be aggregated to obtain poverty
indices. If well-being is measured in the dimensions of income and height, say,
then a person whose income falls below an income poverty line or whose height
falls below a height poverty line could be considered poor. This is a union
definition of multidimensional poverty. By an intersection definition, however,
a person would have to fall below both poverty lines to be considered poor. In
contrast to earlier work, the tests used here are valid for both definitions-or for
any choice of intermediate definitions for which the poverty line in one dimen-
sion is a function of well-being measured in the other dimension.
A second key concept that arises only in the context alf multivariate
poverty comparisons is that, roughly speaking, the correlation between
individual measures of well-being matters. If two populations have the
same univariate distributions for two measures of well-being, but one has
a higher correlation between these measures, then it should not have lower
poverty.6 This is because a person's deprivation in one dimension of well-
being should matter more if the person is also poorer in the other dimension.
The dimensions of well-being are substitutes in the poverty measure. While
this is apparently intuitive, counterexamples are also presented, although the
poverty comparisons are valid only for the case in which the dimensions are
substitutes.
Section I concludes with examples of why the poverty comparisons here are
more general than comparisons of indices such as the United Nations Develop-
ment Programme's human development index (UNDP 1990) and comparisons
that consider each dimension of well-being independently.
Section 11 applies these methods to spatial poverty comparisons in Ghana,
Madagascar, and Uganda, comparing poverty across regions and areas (urban
and rural) in the dimensions of household expenditures per capita and nutri-
tional status for children under the age of 5. Univariate comparisons based on
expenditures or nutritional status alone almost always show greater poverty in
rural areas in any one region than in urban areas in any other region. Bivariate
comparisons, however, are less likely to draw this conclusion lor a variety of
reasons. For this particular application, all of the interesting deviations from the
generally acceptedconclusion that poverty is higher in rural areas result from the
fact that the correlation between these two dimensions of well-being is often
higher in urban areas.
Previous work on multidimensional poverty comparisons has ignored sam-
pling variability, yet this is fundamental if the study of multidimensional poverty
comparisons is to have any practical application. The poverty comparisons here
are all statistical, using consistent, distribution-free estimators of the sampling
distributions of the statistics of each poverty comparison.
6. Bourguignon and Chakravarty (2003, p. 31) refer to this as a "correlation increasing switch" and
discuss it in detail. It is closely related to Tsui's (1999)concept of correlation increasing majorization.
94 THE W O R L D BANK E C O N O M I C REVIEW, VOL. 20, N O . I
This section discusses the data and provides an intuitive presentation of multi-
dimensional poverty comparisons.
Data
The data for this study come from the 1988 Ghana Living Standards Survey, the
1993 National Household Survey (Enqu2tePermanente aupr2s des Mtnages) in
Madagascar, and the 1999'~ationalHousehold Survey in Uganda. All are
nationally representative multipurpose household surveys.
The first measure of well-being is household expenditures per capita, the
standard variable for empirical poverty analysis in developing economies. The
second is children's height-for-age z score (HAZ) that measures how a child's
height compares with the median of the World Health Organization reference
sample of healthy children (WHO 1983). In particular, the z scores standardize a
child's height by age and gender as (xi- x ~ ~ ~ ~ ~ where~xi is a/ child's height,
) u ~ ,
x,,dian the median height of children in a healthy and well-nourished reference
population of the same age and gender, and uxthe standard deviation from the
mean of the reference population. Thus, the z-score measures the number of
standard deviations that a child's height is above or below the median for a
reference population of healthy children of the same age and gender.
The nutrition literature includes a wealth of studies showing that in poor
countries children's height is a particularly good summary measure of children's
general health status (Cole and Parkin 1977; Mosley and Chen 1984; WHO
1995). As summarized by Beaton and others (1990, p. 2), growth failure is
"the best general proxy for constraints to human welfare of the poorest, includ-
ing dietary inadequacy, infectious diseases and other environmental health
risks." They go on to point out that the usefulness of stature is that it captures
the "multiple dimensions of individual health and development and their socio-
economic and environmental determinants." In addition, HAZ is an interesting
variable to consider with expenditures per capita because the two are, surpris-
ingly, not highly correlated, so that they capture different dimensions of well-
being (Haddad and others 2003).~
Univariate Poverty Dominance Methods
The theoretical and statistical bases for the methods used here are developed in
Duclos, Sahn, and Younger (2003). This section provides only an intuitive
presentation; the formal argument is presented in the appendix. Even though
the goal is to make multidimensional poverty comparisons, it is easier to grasp
the intuition with a unidimensional example.
7. Pradhan, Sahn, and Younger (2003)give a more thorough defense of using children's height as a
welfare measure.
Duclos, Sahn, and Younger 95
Consider the question: Is poverty greater in urban or rural areas? The dom-
inance approach to poverty analysis addresses this question by making poverty
comparisons that are valid for a wide range of poverty lines and a broad class
of poverty measures. Figure 1 displays the cumulative density functions-or
distribution functions-for real household expenditures per capita in urban
and rural areas of Uganda in 1999. If the values on the x axis are thought of
as potential poverty lines-the amount that a household has to spend per capita
in order not to be poor-then the corresponding value on the y axis would be the
headcount poverty rate-the share of people whose expenditure is below that
particular poverty line. Note that this particular cumulative density function is
sometimes called a poverty incidence curve. The graph makes clear that no
matter which poverty line one chooses, the headcount poverty index (the share
of the sannple that is poor) will always be lower for urban areas than for rural.
Thus, this sort of poverty connparison is robust to the choice of a poverty line.
What is less obvious is that this type of comparison also permits drawing
conclusions about poverty according to a very broad class of poverty measures.
In particular, if the poverty incidence curve for one sample is everywhere below
the poverty incidence curve for another sample over a bottom range of poverty
lines, then poverty will be lower in the first sample for all those poverty lines and
for all additive poverty measures that obey two conditions: they are nondecreas-
ing and anonymous. Nondecreasing nneans that if any one person's income
increases, the poverty measure cannot increase as well. Anonymous means that
FIGURE 1. Poverty Incidence Curves, Urban and Rural Areas of Uganda 1999
".".
7.5 8.0 8.5 9.0 9.5 10.0
Log ofhousehold expendituresper capita
Source: Authors' analysis based on data from the Uganda 1999 National Household Survey.
96 T H E W O R L D BANK E C O N O M I C REVIEW, VOL. 20, N O . I
it does not matter which person occupies which position or rank in the income
distribution. It is helpful to denote as 111the class of all poverty measures that
have these characteristics. 111includes virtually every standard poverty measure.
It should be clear that the nondecreasing and anonymous characteristics of the
class 111are entirely unobjectionable. Additivity is perhaps less benign, but it is a
standard feature of the poverty measures because it allows subgroup decomposi-
tion (Foster, Greer, and Thorbecke 1984).
Comparing cumulative density curves as in figure 1 thus enables making a
very general statement about poverty in urban and rural Uganda: for any reason-
able poverty line and for the class of poverty measures 111, poverty is lower in
urban areas than in rural areas. This is called first-order poverty dominance. The
generality of such conclusions makes poverty dominance methods attractive.
However, such generality comes at a cost. If the cumulative density functions
cross one or more times, there is no clear ordering-it cannot be said whether
poverty is lower in one group or the other.
There are two ways to deal with this problem, both reasonably general. First,
it is possible to conclude that poverty is lower in one sample than in another for
the same large class of poverty measures, but only for poverty lines up to the first
point at which the cumulative density functions cross (for a recent treatment of
this, see Duclos and Makdissi 2005).If reasonable people agree that this crossing
point is at a level of well-being safely beyond any sensible poverty line, this
conclusion may be sufficient. Second, it is possible to make comparisons over a
smaller class of poverty measures. For example, if the condition is added that the
poverty measure respects the Pigou-Dalton transfer principle,8 it turns out that
the areas under the crossing poverty incidence curves can be compared. If the
area under one curve is less than the area under another for a bottom range of
reasonable poverty lines, poverty will be lower for the first sample for all
additive poverty measures that are nondecreasing, are anonymous, and obey
the Pigou-Dalton transfer principle. This is called second-order poverty dom-
inance, and the associated class of poverty measures is called l12. While not as
general as first-order dominance, it is still a quite general concl~sion.~
Biuariate Poverty Dominance Methods
Bivariate poverty dominance comparisons extend the univariate methods dis-
cussed above. If there are two measures of well-being rather than one, figure 1
becomes a three-dimensional graph, with one measure of well-being on the
x axis, a second on the y axis, and the bivariate cumulative density function
8. The Pigou-Dalton transfer principle says that a marginal transfer from a richer person to a poorer
person should decrease (ornot increase)the poverty measure. Again, this seems entirely sensible, but note
that it does not work for the headcount whenever a richer person located initially just above the poverty
line falls below the poverty 11nebecause of the transfer to the poorer person.
9. If second-order poverty dominance cannot be established, it is possible to integrate once again and
check for poverty dominance for a still smaller class of poverty indices and so on. See Zheng (2000)and
Davidson and Duclos (2000)for more detailed discussions.
Duclos, Sahn, and Younger 97
FIGURE 2. BidimensionalPoverty Dominance Surface
Univariatecumulativedensity
*functionfor household
expenditureper capita
Log household
expenditure per
capita
on the z axis (vertical),as in figure 2. The bivariate cumulative density function
is now a surface rather than a line, and one cumulative density function surface
is compared with another, just as in figure 1. If one such surface is everywhere
below another, poverty in the first sample is lower than poverty in the second
sample for a broad class of poverty measures, just as in the univariate case. It is
also useful to note that univariate poverty incidence curves are the marginal
cumulative densities in the picture found at the extreme edges of the bivariate
surface.
That class, now called II1'' to indicate that it is first order in both dimensions
of well-being, has characteristics analogous to those of the univariate case-
additive, nondecreasing in each dimension, and anonymous-and one more: the
two dimensionsof well-beingmust be substitutes (or more precisely,must not be
complements) in the poverty measure. Roughly, this means that an increase of
well-being in one dimension should have a greater effect on poverty the lower
the level of well-being in the other dimension. In most cases, this restriction is
sensible: if we are able to improve a child's health, for example, it seemsethically
right that this should reduce overall poverty the most when the child is very poor
in the income dimension. But there are some plausible exceptions. For example,
suppose that only healthy children can learn in school. Then, it might reduce
poverty more to concentrate health improvements on children who are in school
(better-offin the education dimension) because of the complementarity of health
and education.
Practically, it is not easy to plot two surfacessuch as the one in figure2 on the
same graph and to see the differences between them, but the differences can be
plotted directly. If this difference always has the same sign, one or the other of
the samples has lower poverty for a large class IT'.' of poverty measures. If the
surfaces cross, the distributions can be compared at higher orders of dominance,
just as in the univariate case. This can be done in one or both dimensionsof well-
being, and the restrictions on the applicable classes of poverty measures are
similar to the univariate case.
INTERSECTION, UNION, AND INTERMEDIATEPOVERTY DEFI~IONS.In addition to the
extra conditions on the class of poverty indices, multivariate dominance com-
parisons require distinguishing among union, intersection, and intermediate
poverty measures. This can be done with the help of figure 3 that shows the
domain of dominance surfaces-the (x,y)plane. The function Xl(x,y)defines an
"intersection"poverty index: someone can be considered poor only when poor
in both dimensionsx and y and therefore when lying within the dashed rectangle
of figure 3. The function X2(x,y)(the L-shaped dotted line) defines a union
poverty index: someone can be considered poor when poor in either of the two
FIGURE 3. Intersection, Union, and Intermediate Dominance Test Domains
Duclos, Sahn, and Younger 99
dimensions and therefore when lying below or to the right of the dotted line.
Finally, X3(x,y)provides an intermediate approach. Someone car1 be considered
poor even with a y value greater than the poverty line in the y dimension if the x
value is low enough to lie to the left of X3(x,y).
For one sample to have less intersection poverty than another for any poverty
line up to z, and z,, its dominance surface must be below the second sample's
everywhere within an area such as the one defined by Xl(x,y).To have less union
poverty, its surface must be below the second sample's everywhere within an
area such as the one defined by X2(x,y)and, similarly, for intermediate defini-
tions and X3(x,y).The X(x,y)function delimits the domain over which domi-
nance tests are compared. As such, it is comparable to the maximal poverty line
in a univariate comparison.
MULTIVARIATEAND HUMAN DEVELOPMENTINDEX POVERTY COMPARIS~ONS. 3 is Figure
also helpful for understanding the difference between the general multivariate
poverty comparisons used here and comparisons that rely on indices created
with multiple indicators of well-being, the best known of which is the human
development index (WP 1990). An individual-levelindex of the x and y mea-
sures of well-being in figure 3 might be written as
where a, and a, are weights assigned to each variable. This index is now a
univariate measure of well-being and could be used for poverty comparisons
such as those in figure 1.'' The domain of this test for such an index would
follow a ray starting at the origin and extending into the (x,y)plane at an angle
that depends on the relativesize of the weights a, and a,. Testing for dominance
at these points only is clearly less general than testing over the entire area defined
by a X(x,y)function in figure 3.
A comparison of poverty in rural Toliara and urban MahajangdAntsiranana
in Madagascar shows why this generalization of human developmentindex-type
univariate indicesis important. Table 1summarizes the value of the t statistic for
a test of the difference in the two areas' poverty surfaces at a 10 x 10 grid of test
points in the domain of figure 3-the (x,y)plane of that figure. The origin (the
poorest people)is in the lower left corner, and the grid of test points is set at each
decile of the marginal distributions.ll The significantly negative differences are
highlighted in light gray and the significantlypositive differences in dark gray. In
10. The human development index is actually cruder than this, as it first aggregates across individuals
each dimension of well-being to generate a single scalar measure and then constructs a weighted average
of those scalars to generate the index, which is also a scalar. Dutta, Pattanaik, and Xu (2003)discuss the
severe restrictions needed on a social welfare function to justify such an index.
11. In theory, differences in the surfaces should be tested for everywhere,but this is computationally
expensive. In practice, because the surfaces are smoothly increasing functions, it is usually sufficient to test
at a grid of points, as is done here.
100 T H E W O R L D BANK E C O N O M I C REVIEW, V O L . 20, N O . I
TABLE 1. IC1>lDominance Tests for Rural and Urban Areas in Toliara,
Madagascar, 1993 (DifferencesBetween Rural and Urban Dominance Surfaces)
0.000 -4.01 -3.33 -2.84 -2.39 -1.98 -1.63 -1.21 -0.71 0.12 4.85
HAZ
Note: The significantly negative differences are highlighted in light gray and the significantly
positive differences in dark gray. Weights a, and a, are chosen, so that a human development index-
type index of these two dimensions of well-being traces out the diagonal, here highlighted in bold.
Source: Authors' analysis based on data from the Madagascar 1993 National Household Survey
(EnquEtePermanente auprks des Mtnages).
choosing the weightsa, and a, so that a human development index-type index of
these two dimensions of well-being traces out the diagonal of table 1, it can be
concluded that poverty is higher in rural Toliara for a wide range of poverty
lines-up to the 70th percentile-and all poverty measures in the 111 class.
However, another choice of a, and a, that gives more weight to household
expenditures would yield test points on a steeper ray from the origin and thus
imply a significant crossing of the index's poverty incidence curves, yielding no
dominance result. Testing over the entire two-dimensional domain rather than a
single ray within that domain avoids this problem.
MULTIVARIATE AND MULTIPLEUNIVARIATEPOVERTY COMPARISONS. Suppose that a
univariate comparison of expenditures per capita in two samples, as in figure1,
and children's heights in two samples finds that for both variables, one sample
shows lower poverty for all poverty lines and a large class of poverty measures.
Is that not sufficient to conclude that poverty differs in the two samples?
Unfortunately, no.
The complication comes from the "hump" in the middle of the dominance
surface shown in figure 2. How sharply the hump rises depends on the correla-
tion between the two measures of well-being. If they are highly correlated, the
surface rises rapidly in the center and vice versa. Thus, it is possible for one
surface to be lower than another at both extremes (the edges of the surface
farthestfrom the origin) and yet higher in the middle if the correlation between
the welfare variables is higher. (The far edges of each surface integrate out one
variable, and so are the univariate cumulative density functions depicted in
figure 1.) Thus, in this case one surface would have lower univariate
Duclos, Sahn, and Younger 101
0.000 -3.100 -2.450 -1.970 -1.580 -1.220 -0.880 -0.500 -0.010 0.690 5.820
HAZ
-- - --
Note: The significantly negative differences are highlighted in light gray and the significantly
positive differences in dark gray.
Source: Authors' analysis based on data from the Uganda 1999 National Household Survey.
cumulative density functions, and thus lower poverty, for both measures of
well-being independently, but it would not have lower bivariate poverty.
Intuitively, samples with higher correlation of deprivation in multiple dimen-
sions have higher poverty than samples with lower correlation because lower
well-being in one dimension contributes more to poverty if well-being is also
low in the other dimension.12
Consider this example. Univariate poverty is unambiguously higher in the
rural Central region of Uganda than in the urban Eastern region in both
dimensions-the difference between the dominance surfaces at the extreme
top and right edges of table 2 is always positive-yet bivariate poverty is not
unambiguously higher because of the statistically significant reversal of the
dominance surfaces in the interior. Similar comparisons up to third order in
each dimension also find that the dominance surfaces cross for these two
areas.
It is also possible that two samples with different correlations between mea-
sures of well-being have univariate comparisons that are inconclusive-they
cross at the extreme edges of the dominance surfaces-but have bivariate sur-
faces that are different for a large part of the interior of the dolminance surface.
(The sample with lower correlation would have a lower dominance surface.)
This would establish different intersection multivariate poverty even though
either one or both of the univariate comparisons are inconclusive. It could not,
12. "Correlation" is actually overly strict. For instance, a recent literature has emerged on copulas,
namely, functions that link two univariate distributions in ways that are more general than simple linear
correlations but less flexible than the nonparametric distributions here. If these copulas differ for two
groups, even if their correlations between dimensions of well-being are the same, it is still the case that
one-at-a-time univariate dominance results could be reversed with a multivariate comparison.
102 THE WORLD BANK E C O N O M I C REVIEW, VOL. 20, N O . I
TABLE 3. I I Dominance Tests for Rural Central and Urban Northern
~ ~ ~
Regions, Uganda, 1999
0.000 -3.100 -2.450 -1.970 -1.580 -1.220 -0.880 -0.500-0.010 0.690 5.820
HAZ
Note: The significantly negative differences are highlighted in light gray and the significantly
positive differences in dark gray.
Source: Authors' analysis based on data from the 1999 Uganda National Household Survey.
however, establish union poverty dominance, since that requires difference in the
surfaces at the extremes as well as in the middle.
Consider the example for rural Central and urban Northern Uganda
(table 3). There is no statistically significant univariate dominance in the
height-for-age dimension of well-being, and only a limited range of poverty
lines for which poverty differs in the expenditure dimension, but there is a
sizable domain-up to the ninth decile in each dimension-over which pov-
erty is lower in the rural Central region than in the urban Northern region for
all intersection poverty indices in the class. Thus, for many intersection
I 1 2 y 2
and intermediate poverty measures, it can be concluded that the rural Central
region in Uganda is less poor than the urban Northern region, even though
neither univariate comparison is conclusive.
This section applies bivariate dominance tests to spatial poverty comparisons
in Ghana, Madagascar, and Uganda. Poverty, measured by household expen-
ditures per capita and children's HAZ, is compared in urban and rural areas,
nationally and by region.13 The tests produce a large amount of output in the
13. The regions used in Ghana are its standard ecological zones of Coast, Forest, and Savannah. In
Uganda, the four political regions are used: Central, Eastern, Western, and Northern. In Madagascar,
political regions are also used, but because of small sample sizes Fianarantsoa and Toamasina are
combined into one region, as are Mahajanga and Antsiranana. This choice is based on similar agro-
ecological characteristics. In all countries, rural and urban areas in these regions are considered.
Duclos, Sahn, and Younger 103
form of tables, such as table l.14Only summaries of the dominance results are
reported here.''
Table 4 gives descriptive statistics for HAZ and the log of household expendi-
tures per capita, ln(y).As expected, poverty measured by expenditures per capita
and also stunting16is higher in rural areas than in urban areas in each country.
The same is true within each region of each country, except for the Toliara
region in Madagascar, where stunting is higher in urban areas than in rural
areas. In fact, with a few exceptions in Madagascar, both expenditure and height
poverty are lower in urban areas in any region of each country than in rural areas
in any other region in the same country.
In addition to the means and poverty rates, table 4 reports the correlation
between the log of expenditures per capita and HAZ. Note that in Madagascar
and Uganda, expenditures and heightsare more highly correlated in urban areas
than in rural areas, whereas both expenditures and heights tend to be higher in
urban areas. As noted, this combination can cause bivariate poverty compar-
isons to differ from univariate comparisons carried out separately in each
dimension of well-being.l7
The dominance results for tests across urban and rural areas in Ghana,
Madagascar, and Uganda show that for each country as a whole, poverty is
higher in rural areas than in urban areas for each univariate poverty comparison
and for both intersection and union bivariate comparisons. These results are
entirely consistent with virtually every known poverty comparison based on
incomes or expenditures alone-poverty is lower in urban areas.
In the regional comparisons, however, a significant number of exceptions to
this widely held belief emerge, especially for the bivariate comparisons. Ghana
has the fewest of these exceptions, with two of nine urban-rural comparisons
being statistically insignificant for both intersection and union bi,variatepoverty
comparisons.1gIn Uganda, for 4 of 16 intersection and union comparisons, the
null hypothesis of nondominance cannot be rejected and two of these-rural
areas in Eastern and Western regions compared with urban areas in the North-
ern region-actually have somewhat limited domains over which bivariate
poverty is lower in the rural area for intersection poverty measures. In
14. The results are relegated to appendixes, which are available from the authors.
15. The relevant statistics and their asymptotic standard errors can be readily computed using the
software DAD (version 4.4 and higher) that is freely available at www.mimap.ecn.ulava1.ca. The authors
can also provide a GAUSS program that does the same.
16. Stunting is usually defined as an HAZof less than -2.
17. It is difficult to find universal explanations for the empirical correlations between indicators. The
reasons are clearlycontext specific. As an example, expenditures and heights may be more highly correlated
in urban than in rural areas because in urban areas the use of food markets may be prevalent. Purchasing
power would then be better correlated with nutrient intake. In rural areas, nutrient intake is plausibly less
correlated wirh purchasing power and more correlated with the proximity of food producers.
18. In each country, rural areas in each region are compared with urban areas in each region. Since
there are three regions in Ghana, this yields nine comparisons. For Uganda and Madagascar, with four
regions, this yields 16 comparisons.
TAB LE 4. Descriptive Statistics for Poverty and Stunting for Ghana,
Madagascar, and Uganda
Mean
Percent
Region HAZ In(y) Stunted Poor N Correlation ln(y),HAZ
Ghana 1988
Coast -0.98 11.90 0.22 0.41 911 0.15
Rural -1.12 11.76 0.27 0.51 488 0.10
Urban -0.82 12.06 0.16 0.30 423 0.15
Forest -1.38 11.81 0.32 0.46 1,074 0.12
Rural -1.48 11.79 0.35 0.48 793 0.11
Urban -1.10 11.88 0.24 0.39 281 0.10
Savannah -1.30 11.66 0.32 0.55 683 0.11
Rural -1.37 11.63 0.33 0.56 591 0.13
Urban -0.86 11.85 0.23 0.48 92 -0.08
National -1.22 11.80 0.28 0.47 2,668 0.14
Rural -1.35 11.73 0.32 0.51 1,872 0.11
Urban -0.92 11.97 0.19 0.35 796 0.11
Madagascar 1993
Tana -2.24 12.32 0.57 0.73 928 0.26
Rural -2.33 12.26 0.60 0.78 534 0.25
Urban -1.80 12.65 0.40 0.48 394 0.20
FianIToa -2.15 12.26 0.53 0.77 975 0.03
Rural -2.19 12.22 0.54 0.80 705 0.00
Urban -1.74 12.56 0.48 0.56 270 0.17
Mahajangal -1.35 12.62 0.34 0.55 561 -0.02
Antsiranana
Rural -1.32 12.61 0.34 0.56 346 -0.04
Urban -1.44 12.71 0.34 0.50 215 0.14
Toliara -1.91 12.06 0.48 0.78 457 -0.18
Rural -1.82 11.98 0.45 0.82 302 -0.19
Urban -2.36 12.46 0.60 0.57 155 0.02
National -1.97 12.33 0.50 0.71 2,921 0.07
Rural -2.01 12.27 0.51 0.75 1,887 0.05
Urban -1.79 12.61 0.44 0.52 1,034 0.17
Uganda 1999
Central -1.00 8.80 0.25 0.19 1,806 0.07
Rural -1.08 8.65 0.27 0.23 1,390 0.04
Urban -0.77 9.22 0.18 0.08 416 0.03
Eastern -1.22 8.48 0.28 0.38 2,349 0.09
Rural -1.25 8.45 0.28 0.39 2,010 0.06
Urban -0.75 8.99 0.21 0.14 339 0.21
Western -1.42 8.63 0.34 0.28 2,096 0.12
Rural -1.46 8.60 0.35 0.29 1,860 0.07
Urban -0.59 9.35 0.15 0.06 236 0.25
Northern -1.24 8.16 0.30 0.60 1,230 0.09
Rural -1.24 8.13 0.30 0.62 1,008 0.08
Urban -1.23 8.72 0.26 0.19 222 0.36
National -1.22 8.54 0.29 0.35 7,481 0.10
Rural -1.27 8.47 0.30 0.37 6,268 0.06
Urban -0.79 9.15 0.19 0.10 1,213 0.12
Source: Authors' analysis based on data from the 1988 Ghana Living Standards Study sumey,
the Madagascar1993 National Household Survey(EnquCtePermanenteauprks des MCnages),and
the Uganda 1999 National Household Suruey.
Duclos, Sahn, and Younger 105
Madagascar, for 7 of 16 intersection comparisons and 10 of 16 union compar-
isons, the null hypothesis that bivariate poverty is the same in urban and rural
areas cannot be rejected, though none of these reject the null in favor of rural
areas. While it is true that in only a minority of cases are urban areas not found
to have significantly lower poverty, the fact that there are any such cases is
surprising, given the overwhelming number of studies that find lower univariate
poverty in urban areas in all developing economies.
One immediate concern with these results is that the interesting cases are the
ones in which the null hypothesis of nondominance is not rejected, so the results
may be driven by a lack of power in the statistical tests. This concern is
reinforced by the relatively few observations available in some urban areas.
Review of the appendix tables shows, however, that in most cases in which
bivariate dominance is not found, the dominance surfaces actually cross signifi-
cantly.19That is, there are points in the test domain where the rural surface is
significantly above the urban surface and vice versa. Thus, the lack of bivariate
dominance is typically not due to a lack of power.
To gain a better understanding of how bivariate and univariate dominance
methods can differ, we classified the results into five types. Type 1 has dom-
inance (usuallyfirst order) for both univariate comparisons and for intersection
and union bivariate comparisons. This is the most common result, accounting
for 25 of the 41 comparisons. This is also the least interesting type of result for
the methods applied here. Why bother with the more complicated bivariate
comparisons if, in the end, they produce the same results as simpler univariate
dominance tests or even scalar comparisons?
Type 2 occurs when neither the univariate nor the bivariate method finds
dominance. This is equally uninteresting for the methods used here. There is only
one such case for urban and rural MahajangaIAntsiranana regon in Madagascar.
Type 3 is a case in which urban areas dominate rural areas for both univariate
comparisons but not for the bivariate comparisons. There are six of these cases.
There is also one case, rural MahajangaIAntsiranana compared with urban
Toliara, in which the rural area dominates on both univariate comparisons but
not on the bivariate comparisons. For cases in which the bivariate comparisons
are inconsistent with the univariate comparisons, a type 3 result is the most
common. The bivariate comparisons are more demanding than univariate com-
parisons, so it makes sense that they reject the null hypothesis o~fnondominance
less often and this happens in five of the sevencases. In two cases, both involving
urban areas in the Northern region of Uganda, the dominance result is actually
reversed for intersection poverty measuresover a limited domain. This is surpris-
ing, but understandable considering the very high correlation (0.36) between
expenditures and heights in urban Northern region compared with rural Wes-
tern region (0.07) and Eastern region (0.06).
19. The results are relegated to appendixes, which are available from the authors.
Type 4 occurs when the univariate results are contradictory in the sense that
univariatedominanceis found in one dimension but not in the other. There are six
such occurrences, and in all but one the urban area dominates in one dimension,
usually expenditures, although in one case, rural Central compared with urban
Northern region in Uganda, the rural area dominates, albeit only for the I I ~ class.
Of these six cases, intersection dominanceis found for four bivariate tests. That is,
the bivariatetests are able to "resolve"the conflicting univariate resultsfor at least
some classes of poverty measures2' and areas of poverty lines.
Type 5 is similar to type 4 except that the contradictory univariate results are
statistically significant in each univariate comparison. There are only two of
these cases, rural compared with urban Toliara and rural Coast compared with
urban Forest in Ghana. Unlike the type 4 results, in neither case are any of the
bivariate poverty comparisons statistically significant, so the bivariate compar-
isons cannot resolve the univariate conflict.
Overall, sufficient evidence has not been amassed to overturn the standard
presumption that poverty is lower in urban than in rural areas, but enough of the
results are at odds with this idea to introduce doubt. Furthermore, the reasons
that this is not found for bivariate poverty comparisons vary. For the type 4 and
5 cases, no univariate dominance is found in one dimension or another and the
bivariate results follow from that. But this is relativelyrare, and in about half of
the cases the bivariate tests for intersection poverty measures do find that
poverty is lower in urban areas despite the contradictory univariate results.
Most of the differences, though, come from the fact that the two measures of
well-being are often more highlycorrelated in urban areas than in rural areas. As
noted, this correlation causes the poverty incidence surface to rise more rapidly
near the origin of the distribution, raising it above the rural surface in the center
even though it is belowit at the extremes, where the univariate poverty incidence
curves lie. In most cases, this gives results in which an urban area dominates a
rural area in each dimension individually, but not jointly, because multiple
deprivation is more common in urban areas. There are two cases, however, in
which the dominance is reversed, so that for some intersection poverty measures
the rural area dominates the urban area.
This article used bivariate stochastic dominance techniques to compare poverty
in urban and rural areas in three African countries, measuring poverty in terms
of expenditure per capita and children's standardized heights, a good measure of
children's health status. The comparisons are shown to be more general than
either a comparison of a human development index-type index or one-at-a-time
20. As noted in the methods discussion, bivariate dominance for union poverty measures requires
univariate dominance in each dimension, so it is impossible for this type of result.
Duclos, Sahn, and Younger 107
comparisons of multiple measures of well-being. More important, the article
finds that its more general methods are at odds with simpler univariate poverty
comparisons in a nontrivial number of cases.
Expenditure-based urban-rural poverty comparisons almost always find that
rural areas are poorer than urban areas. The results are consistent with that
finding whether univariate or bivariate comparisons are used. However, differ-
ences emerge when urban areas in one region of a country are compared with
rural areas in another region. In several cases, univariate poverty is lower in
urban areas in both dimensions, but bivariate poverty is not. This happens
because the correlation between expenditures per capita and children's heights
is higher in the urban areas, so that urban residents who are expenditure poor
are also more likely to be health poor. This correlation yields a higher density of
observations in the poorest part of the bivariate welfare domain for urban areas,
even though there are fewer observations for urban residents at the lower end of
the density for each individual measure of well-being. Taking such a correlation
into account is important for welfare comparisons because the social cost of
poverty in one dimension, say health, is higher if the person affected is also poor
in the other dimension (expenditures in this case).
It is interesting to note that the share of cases in which urban areas do not
dominate rural is much higher in the bivariate comparisons than in the expendi-
ture- or income-based comparisons in the literature, where poverty is almost
always found to be lower in urban areas. With two exceptions in Madagascar,
however, the urban area in the region where the capital city is located always
has lower poverty than every rural area in both univariate and bivariate compar-
isons. Thus, the doubts raised here apply only to other urban areas in these
countries.
There are other instances in which the bivariate comparisons are at odds with
univariate comparisons. Perhaps the most interesting are cases in which univari-
ate results are inconclusive because one or the other univariate comparison is
inconclusive, yet the bivariate results find dominance for a large domain of
intersection poverty indices. This arises in about 10 percent of the examples
and occurs again when the correlation between expenditures per capita and
children's heights differs significantly across areas. These results are interesting
because they show that bivariate comparisons may provide statistically signifi-
cant results when univariate comparisons do not.
The finding that bivariate results often hffer from the standard perception of
greater ruralpovertytypicallyoccursnot becausechildrenaretallerinruralareas, but
rather because the correlation between expenditures and heights is lower there than
in urban areas. This, however, is based on only three countries. Pursuing similar
researchinothercountrieswillyieldinsightintowhether theseresultsare anomalous.
Why this should be is also an interesting question for future research. But a clear
implication of these resultsfor researchers and policymakersinterested in multiple
dimensions of poverty is that, at a minimum, one should check the correlations
between measures of well-being in the groups of interest. Large differences in
these correlationsmay lead to unexpectedmultivariate dominancecomparisons.
The following is based on the companion paper, Duclos, Sahn, and Younger
(2003).
Making Poverty Comparisons with Multiple Indicators of Well-Being
For expositional simplicity, the focus is on the case of two dimensions of
individual well-being. Let x and y be two such indicators. Assuming differentia-
bility, denote by
a summary indicator of individual well-being, analogous to but not necessarily the
sameasa utilityfunction.Notethat thederivativeconditionsinequation (A.l)simply
mean that different indicatorscan each contribute to overall well-being. Assume that
an unknown povertyfrontierseparates the poorfrom the rich, defined implicitly by a
locus of the form A(x,y) = 0 and analogous to the usual downward sloping
indifference curves on the (x,y)space. The set of the poor is then obtained as:
Letting the joint distribution of x and y be denoted by F(x,y), assume for
simplicity that the multidimensional poverty indices are additive across indivi-
duals and define such indices by P(A):
where ~ ( xy;,A) is the contribution to poverty of an individual with well-being
indicators x and y:
2 0 ifA(x,y) 50
= 0 otherwise.
Here, TC is the weight that the poverty measure attaches to someone inside the
poverty frontier. By the focus axiom, it has to be 0 for those outside the poverty
frontier. A bidimensionalstochastic dominance surface can then be defined as:
This function looks like a two-dimensional generalization of the Foster-Greer-
Thorbecke index and can also be interpreted as such. The poverty comparisons
Duclos, Sahn, and Younger 109
here make use of orders of dominance, s, in the x and s, in the y dimensions,
which correspond respectively to sx = a; + 1and s, = ay 1.
+
Assume that the general poverty index in equation (A.3)is left differentiable
with respect to x and y over the set A(A), up to the relevant orders of dominance,
s, for derivatives with respect to x and sy for derivatives with respect to y.
Denote by 7$ the first derivative of ~ ( xy;,A) with respect to x, by ./TY the first
derivative of T(X,y;A) with respect to y, and by ?rxy the derivative of ~ ( x , yA);
..1 1
with respect to x and to y. The following class II ' (A*)of bidimensional poverty
indices can be defined as:
A(A)c A(A*)
~ ( xy;,A) = 0 whenever A(x,y)= 0
(A.6) ]nlll(A*)=
.ir"IOandd'<O vx,Y
f l y2 0 'dx,Y
The first line on the right of equation (A.6) defines the largest poverty set to
which the poor must belong: the poverty set covered by the P(A) indices should
lie within the maximal set A(A*). The second line assumes that the poverty
indices are continuous along the poverty frontier. The third line says that indices
that are members of ]nl'lare weakly decreasing in x and in y. The last line
assumes that the marginal poverty benefit of an increase in either x or y
decreases with the value of the other variable.
Denote by A F = FA - FB the difference between a function F for A and for I?.
The class of indices defined in equation (A.6) then gives rise to the following
theorem:
Theorem 1
AP(A) > 0, 'dP(A)E fil" (A*),
(A.7) if A P ~ ) ~ ( X>,0,)
~ 'd(x,y)E A(A*).
Proof: Denote the points on the outer poverty frontier X*(X,Y)= 0 as z,(~)
for a point above y and Z,(X) for a point above x. The derivative conditions in
-
equation (A.l) imply that zil)(y)5 0 and zp)(x)I 0, where the superscript 1
indicates the first-order derivative of the function with respect to its argument.
Note that the inverse of z,(y) is simply zy(x):x z,(z,(x)). Next, equation
(A.3)is integrated by parts with respect to x,over an interval of y ranging from
0 to z,. This gives:
To integrate by arts with respect to y the second term, define a general
Ph(x,y)l(x,y)dxand
function K(y)=f y note that:
110 THE WORLD BANK E C O N O M I C REVIEW, VOL. 20, N O . I
g(y)ah(x,y)
l(x,y)dx
+ S o -&-
Reordering equation (A.9) and integrating it from 0 to c yields:
(A.10)
Now replace in equation (A.10) c by zy,g(y) by zx(yl, h(x,y) by ?(x, y;A*),
l(x,y) by P(x,y), and K(y) by its definition K ( ~=) J,, h(x,y) l(x,y)dx. This
gives:
S zz (ZY)
P ( z ~ ( ~zy)l=- o
) rr'(x,zy;A*)PolO(x,zy)dx
r(')
+ 1;
rr'Y(x;y;A*)P ~ ~ ~Y)dx,dy.
( x
0
For the sufficiency of equation (A.7), recall that ZL')(~) 0, ? 5 0, and
<
?Y > 0, with strict inequalities for either of these inequalities over at
least some inner ranges of x and y. Hence, if APOlO(x,y) > 0 for all y E [O, zy]
and for all x E [O,zx(y)](that is, for all (x,y) E A(A*)), it must be that
AP(A*) > 0 for all of the indices that use the poverty set A(A*) and that
obey the first two lines of conditions in equation (A.6). But note that
for other poverty sets A(A) c A(A*), the relevant sufficient conditions are
only a subset of those for A(A*). The sufficiency part of theorem 1 thus
follows.
For the necessity part, assume that APO,O(x,y) 5 0 for an area defined over
x E [c;, c,+] and y E [c;, c;], with c,+ <zy and c,+ < zx(y).Then any of the
,.li
poverty indices in I3 ' (A*) for which ?Y < 0 over that area, ?Y = 0 outside
that area, and for which ?(x,zy; A*)= ?(zX(y), y;A*) = 0, will indicate that
AP < 0. Equation (A.7) is thus also a necessary condition for the ordering
specified in theorem 1.
Duclos,Sahn,and Younger 111
Note that similar proofs are possible for dominance comparisons at higher
orders (Duclos, Sahn, and Younger 2003).
Estimation and Inference
Suppose a random sample of N independently and identically distributed
observations drawn from the joint distribution of x and y. 'These observa-
tions of xL and YL , drawn from a population L, can be written as
(x:,y!), i=1,...,N. A natural estimator of the bidimensional dominance
surfaces Paxlay(z,, Z, ) is then:
where F denotes the empirical joint distribution function, I(.) is an indicator
function equal to 1 when its argument is true and 0 otherwise, and
x+ = max(0, x).This give rise to theorem 2:
Theorem 2
Let the joint population moments of order 2 of (z,- yA): (z, - xA)>nd
(z, - yB): (z, - xB)? be finite. Then N112(pA (z,, z,)
az,ay
-P?'~'(Z,, zy)) and
A
~ 1 1 2p;lCYy
( (e,-~2'"' z,)) are asymptotically normal with mean 0,
zy) (z,,
with asymptotic covariance structure given by (L,M = A,B):
lim ~ c o v ( ~ (: G~, z , ) , P ~(zx,zy))=E ((zy- (z,-xL)y ($ -yM)?
' ' ~ yL)y
N'cc
Proof: For each distribution, the existence of the appropriate population
moments of order 1 permits application of the law of large numbers to
&,ay
equation (A.121, thus showing that PK A (zX,zy)is a consistent estimator of
~ ~ ' " ' ( z , , z Given also the existence of the population moments of order 2,
~ ) .
the central limit theorem shows that the estimator in equation (A.12) is
root-N consistent and asymptotically normal with asymptotic covariance
matrix given by equation (A.13). When the samples are dependent, the co-
variance between the estimator for A and for B is also provided by equation
(A.13).
112 T H E W O R L D BANK E C O N O M I C REVIEW, VOL. 20, N O . I
Atkinson, A. B. 1987. "On the Measurement of Poverty." Econometrica 55(4):749-64.
. 2003. "Multidimensional Deprivation: Contrasting Social Welfare and Counting Approaches."
Journal of Economic Inequality 1(1):51-65.
Atkinson, A. B. , and F. Bourguignon. 1982. "The Comparison of Multi-Dimensional Distributions of
Economic Status." In A. B. Atkinson, ed., Social Justice and Public Policy. London: Harvester
Wheatsheaf.
. 1987. "Income Distribution and Differences in Needs." In G. R. Feiwel, ed., Arrow and the
Foundations of the Theory of Economic Policy. New York: New York Press.
Beaton, G.,A. Kelly, J. Kevany, R. Martorell, and J. Mason. 1990. "Appropriate Uses of Anthropometric
Indices in Children: A Report Based on an ACCISCN Workshop." ACCISCN State-of-the-Art Series,
Nutrition Policy Discussion Paper 7. United Nations, Administrative Committee on Coordination1
Subcommittee on Nutrition, New York.
Bourguignon, F. , and S. R. Chakravarty. 2002. Multi-dimensional Poverty Orderings. Paris: DELTA.
-. 2003. "The Measurement of Multidimensional Poverty." Journal of Economic Inequality
1(1):25-49.
Cole, T. J. , and J. M. Parkin. 1977. "Infection and Its Effect on Growth of Young Children: A
Comparison of the Gambia and Uganda." Transactions of the Royal Society of Tropical Medicine
and Hygiene 71:196-8.
Crawford, I. A. 1999. "Nonparametric Tests of Stochastic Dominance in BivariateDistributions, with an
Application to UK." Discussion Papers in Economics 99/07. London, UK. University College.
Davidson, R. , and J.-Y. Duclos. 2000. "Statistical Inference for Stochastic Dominance and for the
Measurement of Poverty and Inequality." Econometrica 68(6):1435-65.
Duclos, J.-Y. , D. Sahn, and S. D. Younger. 2003. "Robust Multidimensional Poverty Comparisons."
Working Paper 98. Ithaca, NY: Food and Nutrition Policy Program, Cornell University.
Duclos, Jean-Yves, and Paul Makdissi. 2005. "Sequential Stochastic Dominance and the Robustness of
Poverty Orderings." Review of Income and Wealth 51(1):63-88.
Dutta, I. , P. K. Pattanaik, and Y. Xu. 2003. "On Measuring Deprivation and the Standard of Living in a
Multidimensional Framework on the Basis of AggregateData." Economics 70(278):197-221.
Foster, J. E. 1984. "On Economic Poverty: A Survey of Aggregate Measures." In R. L. Basmann, and
G. F. Rhodes, eds., Advances in Econometrics, Vol. 3. Greenwich, CT: JAI Press.
Foster, J. E. ,J. Greer, and E. Thorbecke. 1984. "A Class of Decomposable Poverty Measures." Econo-
metric~52(3):761-76.
Foster, J. E. , and A. F. Shorrocks. 1988a. "Inequality and Poverty Orderings." European Economic
Review 32(2/3):654-62.
. 1988b."PovertyOrderings." Econometrica 56(1):173-7.
-. 1988c. "Poverty Orderings and Welfare Dominance." Social Choice Welfare 5(U3):179-98.
Haddad, L. ,H. Alderman, S. Appleton, L. Song, and Y. Yohannes. 2003. "Reducing Child Malnutrition:
How Far Does Income Growth Take Us?" World Bank Economic Review 17(1):107-31.
Mosley, W. H. , and L. C. Chen. 1984. "AnAnalytical Framework for the Study of Child Survival in
Developing Countries." Population and Development Review lO(Suppl.):25-45.
Pradhan, M. ,D. E. Sahn, and S. D. Younger. 2003. "DecomposingWorld Health Inequality."Jounal of
Health Economics 22(2):271-93.
Tsui, K. 1999. "Multidimensional Inequality and Multidimensional Generalized Entropy Measures: An
Axiomatic Approach." Social Choice and Welfare 16:145-48.
.2002. "Multidimensional Poverty Indices." Social Choiceand Welfare 19(1):69-93.
UNDP (UnitedNations DevelopmentProgramme).1990. Human Development Report. New York: Oxford
University Press.
Duclos, Sahn, and Younger 113
omro (WorldHealth Organization).1983. "Measuring Change in Nutritional Status: Guidelinesfor Asses-
sing the Nutritional Impact of Supplementary Feeding Programmes for Vulnerable Groups." Geneva.
.1995. "An Evaluation of Infant Growth: The Use and Interpretation of Anthropometry in
Infants." Bulletin of the World Health Organization 73:165-74.
Zheng, B. 2000. "Poverty Orderings."Journal of Economic Surveys 14(4):427-66.
The Impact of Regulatory Governance
and Privatization on Electricity Industry G.eneration
Capacity in Developing Economies
John Cubbinand JonStern
This article assesses whether a regulatory law and higher quality regulatory governance
are associated with superior outcomes in the electricity industry. The analysis, for 28
developing economies over 1980-2001, draws on theoretical and empirical work on the
impact of telecommunications regulators in developing economies. Controlling for
privatization and competition and allowing for country-specific fixed effects, both
regulatory law and higher quality regulatory governance are positively ar~dsignificantly
associated with higher per capita generation capacity. This positive impact increasesfor
more than 10 years, as experience develops and regulatory reputation grows. The
results are robust to estimating alternative dynamic specifications (including error
correction models), to inclusion of economy governance political risk indicators, and
to controlling for possible endogeneity biases. The article concludes with a short
discussion of causality in panel data modeling of governance models and the policy
implications for regulatory reform.
Over the past 10-15 years, much attention has gone to the role of institutions in
economic growth-in large part to determine the economic policy priorities for
developingfunctioning market economiesin Central and Eastern Europe and the
former Soviet Union and fostering economic growth in lagging regions, such as
sub-Saharan Africa.
In parallel, and partly in response, there have been major explorations of the
role of institutions in market economies. There has also been a substantial
empirical literature on the relative roles of institutions, policy, geography, and
trade openness on growth across economies. This literature places considerable
weight on institutional quality as a major determinant of variations in long-term
g0wth.l In particular, Rodrik (2003, p. 25) argues that there is a requirement
John Cubbin is a professor of economicsat City University in London; his email address is j.s.cubbin@
city.ac.uk. Jon Stern is a regulation initiative associate at London Business School; his email address is
jstern@london.edu.Work on this article was supported by the researchprogram on Industrial Organization
Policy for Development at the Development Research Group of the World Bank, under the direction of
Ioannis Kessides.The authors are grateful for helpful comments from seminar participants at the University
of Cambridge and City University as well as from the editor and three anonymous referees.
1. See Rodrik, Subramanian, and Trebbi (2002)for a recent survey of the literature on studies of
cross-country growth performance.
THE WORLD BANK ECONOMIC REVIEW,VOL. 20, NO. 1, pp. 115-141 doi:10.1093/wber/lhj004
Advance Access publication March 23, 2006
O The Author 2006. Published by Oxford University Press on behalf of the International
Bankfor Reconstructionand DevelopmentITHE WORLD BANK.All rights reserved. For permissions,
please e-mail: journals.permissiom@oxfordjomals.org.
for a "cumulative process of institution building to ensure that growth does not
run out of steam and that the economy remains resilient to shocks."
These arguments apply with extra force to utility service industries-not just
because they are highly capital intensive but also because most of their assets are
very long lived and (in economic terms) sunk. So, an effective institutional
framework is essential to sustain growth in output, efficiency, and capacity for
commercialized utility service industries, such as electricity, telecommunica-
tions, and water, particularly if they have significant private investment (physical
or financial).
The standard institutional solution to handle these infrastructure industry
issues is to introduce an independent regulatory agency, operating within a
clearly defined legal f r a m e ~ o r kThe . ~ agency is intended to provide the high-
quality institution that permits and fosters sustained growth in capacity and
efficiency in the utility service industries, particularly the network elements. So,
whether country X has a high- or a low-quality institution is determined pri-
marily by the quality of governance of the regulatory agency (conditional on the
quality of governance for the economy as a whole). Developing economies with
high-quality regulatory agencies (asmeasured by regulatory governance)should
attract more sustained investment into their utility service industries and at a
lower cost of capital. The regulated utilities should also have higher efficiency
and growth rates.
The perspective just outlined is at the heart of the recent literature on reg-
- -
ulatory governance for utility service industries, particularly focusing on devel-
oping and transition economies. It is set out in Levy and Spiller (1994)-which
draws explicitly on North (1990)-andin subsequent papers (see,among others,
Smith 1997; Stern and Holder 1999; No11 2000). But there have also been
apparently disappointing outcomes of regulatory (and electricity) reforms. The
many case studies, though illuminating, do not allow reliable generalizations.
Only recently has there been formal statistical or econometric testing of the
view-and policy-that better regulatory governance increases investment and
efficiency in the electricity industry.
This article provides an econometric analysis of the relationship between the
quality of regulatory governanceand the level of generation capacity per capita for
electricity supply industries in 28 African, Asian, Caribbean, and Latin American
countries over 1980-2001, controlling for privatization and competition.
2. An independent regulatory agency is not the only way of providing the necessary institutional
support either in theory or in practice (see Domah, Pollitt, and Stern 2002). In addition, an independent
regulator may be combined with a high or low degreeof relianceon contracts and courts. A major issue is
whether low-income developing economies have the human and other resources to sustain independent
regulatory agencies, particularly regulatory agencies with a significant degree of discretion. Nevertheless,
an independent regulatory agency has become the standard recommended solution to the private invest-
ment problem for utilities in the same way as an independent central bank has become the standard
recommended solution to handle commitment and time inconsistency problems in monetary policy (see
Stern and Cubbin 2004).
Cubbin and Stern 317
This article first outlines the underlying economic issues and the main institu-
tional design considerations and related recent research. It then outlines the
modeling approach. The main econometric results from static and dynamic
models are examined, followed by a discussion of endogeneity and causality
issues.3
I. UNDERLYING ECONOMIC ISSUES, INSTITUTIONAL DESIGN,
AND IMPLICATIONS FOR EMPIRICAL ANALYSIS
The main focus here is the inability of governmentsto make credible and binding
commitments about utility pricing to sustain private investment while retaining
decisionmaking powers.4 The discussion concentrates on commercialized utili-
ties facing genuine budget constraints, particularly where private investment and
private finance are important, with a focus on regulatory governance (autonomy
and accountability) rather than regulatory content (pricesand investments).The
underlyingeconomic issue for utility regulation-as for monetary policy-is that
governments, particularly at certain times, have a strong incentive to behave in a
shortsighted and populist manner that reduces welfare over a medium- to long-
term period.
Output Measures for Utility Regulatory Agencies
For utility service industries, there are two main output measures for utility
regulation:
The level and rate of growth of technical efficiency and productivity (and
the quality of service).
The level of capacity.
This article focuses on capacity in developingeconomies,testing the key policy
objectiveof the World Bank and many countries in the sample: that significantly
higher investment (and private investment) is the most important reason for
promoting independent regulatory agencies in electricity and similar utility
service industries (WorldBank 1994, chapter 3).Accordingto this view, follow-
ing the establishment of a regulatory agency, there should be:
Sizable increases in investment flows (domestic and foreign) to developing
economy electricity industries.
Larger increases with higher quality regulatory governance.
Larger impacts as the regulatory agency gains experience and reputation.
3. See Cubbin and Stern (2004)for a much fuller version of the study, particularly the data and tables
of results.
4. The problems for developing economy governments in making credible commitments to support
new investment in the presence of major fixed costs arise in other contexts beside utility regulation. A
good example is export taxes for exportable cash crops (seeMcMillan 2001).We are grateful to the editor
for this observation.
Previous Literature
The empirical work here adopts and extends the fixed-effects panel data model-
ing used in the literature on the impact of regulation on telecommunications
outcomes (see,for instance, Fink, Mattoo, and Rathindran 2002; Wallsten 2002;
Gutierrez 2003).
The approach of Gutierrez (2003) is particularly relevant. He constructs a
regulatory governance index for his sample of 22 Latin American and Caribbean
countries. A seven-element index (derived from the Stern-Holder typology) is
calculated from examining each country's telecommunications laws and changes
in the laws. (Thisarticle's model for electricityoutcomes adopts a similar approach
and uses a "snapshot" four-element index as one regulatory variable.) Gutierrez
(2003)finds statisticallysignificant positive direct effects of his regulatory index on
both teIedensity and efficiency, in static and dynamic models and after testing for
the endogeneity of regulation. The estimated effect of a 1 percentagepoint increase
in the index on mainlines per 100 inhabitants varies somewhat depending on the
precise model specification but is, in general, on the order of 20 percent.
For electricity,there are only a few (verypreliminary) empirical studies of the
impact of regulation, such as Zhang, Kirkpatrick, and Parker (2002)and part of
Pargal (2003).They concentrate on generation capacity, as does this article, but
find only weak effects (if any) of regulation. Their studies also have major
problems in disentangling the effects of regulation from those of privatization
and liberalization. Drawing on a 2001 study by Domah (seeDomah, Pollitt, and
Stern 2002 for details),' this study had access to better data on regulatory
governance and its variation across countries, but again data constraints con-
fined the estimates to capacity models for generation rather than transmission,
distribution, sales, or commercial 10sses.~
Regulatoryissues are, of course, only one aspect of electricity industry reform.
For a comprehensive discussion of electricity reform in developing economies,
see Jamasb and others (2005).
The modeling work reported here is concerned with whether better regulatorygov-
ernancein developingeconomiesincreasesthe rated generationcapacityper capita.
Economic Rationale
For the electricity industry in developing economies, explicit regulation focuses
on providing sufficient supplies, and that typically means increasing investment
5. The authors are gratefulto PreetumDomah for permissionto use the informationfrom hissurveyhere.
6. Cubbin and Stern (2004)report estimates of the impact of regulation on technical losses.
Cuhbin and Stern 119
and capacity. In some cases, this has been done by harnessing private ownership
and domestic or foreign private investment. In others, a workable financial
framework is ~rovidedfor the electricity industry to develop by loosening the
ties with government. A country might enact an electricity law giving various
powers and duties to a ministry or independent regulator. Such changes can also
increase ~ublicinvestment in infrastructure-say, by requiring state-owned
electricity companies to operate more commercially and allowirlg them to use
private debt finance for investment at a reasonable cost of capital.
Investment is encouraged once effectiveregulation supports a workable finan-
cial framework. If the electricityindustry is privatelyowned, the owners have the
prospect of earning a reasonable return on their investment. If publicly owned,
the industry can become independent of tax revenue or continu~allyincreasing
loans. In addition, an effective regulatory framework can encourage the growth
of private investment and private finance within state systems, as in India and
China in recent years.
In an unconstrained market economy, per capita generation capacity adjusts
to demand, which depends on per capita income, the price of electricity, and
such environmental factors as climate. The priceof electricitywill be determined
in part by the efficiency of the sector, with efficiency depending on regulatory
factors and energy sources, such as coal, gas, hydro, and oil. Many developing
economies with a traditional, vertically integrated, and state-ovvned electricity
sector will be constrained not so much by market demand but by the availability
of a continuing subsidy. In electricitymarkets with implicit or explicit subsidies,
capacity constraints arise because of inadequate government revenues for elec-
tricity investment or subsidy payments or insufficient revenue flows to support
viable private investment or commercial debt obligations-or both.
Electricity generation models for unconstrained markets typically find per
capita GDP to be the major determinant of electricity demand (and thus of genera-
tion capacity). Therefore, per capita GDP is included in the model here as well as
other control variables found to be statisticallysignificant in previous studies of
developingeconomy infrastructure industry,such as the share of iindustry in value
added, country debt levels, and country economywide governance indicators.
Variables for electricity privatization and competition are also included.
An effective regulatory framework can be expected to reduce the constraint
on the operation of the market, increasingsupply and moving the outcome closer
to the market equilibrium. The better the governance of the regulator, the
greater the expected increases in capacity and electricity supply.
Model Specification
For developing economies, with supply-constrained electricity, improved regu-
latory governance is expected to raise the equilibrium generation capacity levels.
But the adjustment to the new equilibrium is very likely to take some time to
achieve.
120 THE W O R L D BANK E C O N O M I C REVIEW, VOL. 20, N O . I
This suggests a long-run static model of the followingform, specified below in
panel data format:
LogELCAPPCisthe logof per capita electricitygeneration capacityingigawatts,
a. is a constant term, ai is a time-invariant country-specificfixed effect, GDPPC is
real per capita national income in 1995 U.S. dollar^,^ RegVar comprises one or
more of the regulatory governance variables, X is a vector of other possibly
relevant sectoral and country-level control variables, and vit is an error term.
In all cases, the variables are defined for i = 1,... ,I countries over
t =1,...,T time periods.
The X vector of control variables for this equation might well include domes-
tic fuelhydro source availability and a variety of other country-specific eco-
nomic or institutional variables. But both of these are expected to be captured
largely by the country-specific fixed effects. Similar arguments apply to institu-
tional and country governance effects since country rankings on these indicators
tend to be fairly stable over 10-20 year periods. Also explored is whether either
privatization or competition affects generation capacity growth; both direct and
indirect effects are investigated (suchas interactions between these variables and
the regulatory variables). And on the basis of previous studies of electricity
demand, al would be expected to be close to but probably less than
Equation (1)is a static representation of the model, which provides evidence
on long-run equilibrium effects. Some dynamic error correction models are also
considered, which provide evidence on the adjustment time path and separate
short-run adjustment effects from long--run equilibrium effects.
To ensure that the modeling yieldsestimates of supply responses, the sample is
confined to countries with unsatisfied demand for electricity throughout 1980-
2001-that is, to developing economies in Africa, Asia, and Latin America.
Developed economies and European transition economies are excluded because
both have significant planning margins or unused capacity for some if not all
years of the period.
Data
The sample is of 28 developing economies, for which there is complete (or near
complete) generation capacity data for 1980-2001. This gives a longer panel
than is usually available for such studies, greatly reducing the econometric
problems associated with short panels. But because of some missing observa-
tions, the panel is unbalanced.
Of the 28 countries in the sample, 15 are in Latin America, six in the
Caribbean, five in Africa, and four in Asia. They include large countries (Brazil
7. DP is on an exchange rate basis rather than a purchasing power parity basis.
G
8. See Dahl and Roman (2004)table 5 for a recent survey of electricity demand elasticities.
Cubbin and Stern 121
and India), small countries (Jamaica), middle-income countries (Chile and
Mexico), and poor countries (Ethiopia and Sudan).
The full list of countries and summary regulatory characteristics is in appen-
dix table A.1. Of the 26 regulatory reforms listed, only six are before 1995.
The dependent variable in the regressionsis per capita generation capacity by
country and year, derived from U.S. Energy Information Agency data on gen-
eration capacity by country for 1980-2001 (appendix figure A.1). There are
some significant decreases (Nigeria and Nicaragua 1990-95) as well as large
increases (Paraguay 1985-88). Generation capacity changes tend to be lumpy, so
that the dependent variable does not obviouslyexhibit common olr stable trends.
(Note that the series does not distinguish between publicly and privately owned
generation capacity.)
The key independent variables for this study are the regulatory variables
available-data for each country on the existence (or absence) of:
An electricity (or energy) regulatory law.
An autonomous or ministry regulator.
License fee or government budget regulatory funding.
Free or mandatory civil service pay scales for regulatory staff.
Each is measured by a 0 or 1dummy variable. The dating of the switch from 0
to 1on the appropriate variables (subsequentlymaintained at a c~onstantlevel)is
derived from the date of enactment of a primary electricity reform or regulatory
law (except where other information provided a known, superior alternative).
So, the effect of age of the regulatory agency as well as its existence can be
investigated to estimate alternative measures of the impact of regulation. Given
the time needed to establish a functioning regulatory entity, the start date for
regulation is taken as the year following the enactment of the law.
The regulatory variables in the index are all measures of formal attributes of
regulation. No comparable data are available on the informal, practical qualities
of electricity regulation, such as the transparency and quality of regulatory
processes. The necessary omission of data on these characteristics may lead to
biased estimates and standard error^.^ And unlike Gutierrez (2003),there is no
time dimension on changes in formal governance attributes after the enactment
of the primary electricitylenergyregulatory law.''
The Domah data set, suitable for a preliminary investigation of the impact of
regulation, is not ideal. It suffers from an absence of data on the informal,
practical aspects of regulation, such as the length of tenure of regulatory agency
heads or commissioners.
9. See Stern and Cubbin (2004),pp. 30-32, where preliminary simulation results based on the Stern-
Holder data set suggest that omitting data on the informal, practical aspects of regulation can lead to
coefficients being underestimated by about 5-10 percent and a similar underestimate of t-values.
10. For a fuller description of the data and a range of descriptive statistics, see Cubbin and Stern
(2004),section IV, p. 19.
122 T H E W O R L D B A N K E C O N O M I C REVIEW, VOL. 20, N O . I
Although much of the regulatory activity took place in the second half of the
period, the earlier period is important for establishing benchmark pre-reform
levels of generation capacity and for reducing some of the biasesthat can arise in
the use of short panels. Of the total number of country-sample years, 21 percent
were with an autonomous regulator and 31 percent with an electricity or energy
regulatory law. By the end of the period, only two of the countries had not
enacted an electricity law and there were nine with a ministry regulator operat-
ing under a law.''
A key feature of the regulatory data is that the correlation between the four
regulatory variables is (not surprisingly) very high (table 1). In addition, all
countries with an autonomous regulator had an electricity law, as did all the
countries with license fee funding.12This high collinearity between the regula-
tory variables presents estimation problems, discussed in the next section.
For privatization and competition, the Henisz-Zellner-Guilen (HZG; Henisz,
Zellner, and Guillen 2004) electricity data are used.13The data on privatization
provide information on the year all countries introduced one of the following:
minority privatization of their electricity industries, majority privatization of
their electricity industries, and total privatization of their electricity industries.
The HZG data on competition includea variable for the year private firms were
legally allowed to generate electricity for resale. But this does not necessarily
mean that such electricity sales were important or even took place, and half the
countries in the sample had this attribute over the whole period. More seriously,
this variable provides no information on the market structure of generation or
wholesale electricity purchasing. But it is the only consistently available "com-
petition" variable for developingeconomies over the period.
The other main data source was World Bank data, including the World Bank
World Development Indicators (as for per capita GDP and population) and the
Kaufmann governance indicators.
TABLE 1. Correlation Matrix of Regulatory Variables
Electricity License Fee Autonomous Non-Civil Service
Law Funding Regulator Pay Scales
Electricity law 1.OOO
License fee funding 0.849 1.000
Autonomous regulator 0.783 0.703 1.000
Non-civil service pay scales 0.783 0.551 0.443 1.000
Source: Authors' analysis based on data described in the text.
11. Fora fullerdescriptionof thedata anda rangeof descriptivestatistics,see CubbinandStern2004,
Section 4, p 19.
12. Uruguaywasa partialexception,introducinglicensefeefunding,3yearsbeforeitslawcameintoforce.
13. We are grateful to ProfessorHenisz for permission to use these data.
Cubbin and Stern 123
This section covers results from a static, long-run model, dynamic models, and
endogeneity and causality issues.
Econometric Results for Models of Generation Capacity: Static Model
An ordinary least squares (OLS) equation was estimated as a baseline. However,
all the equations reported here are modeled using a fixed-effects estimator.
Moving from OLS to a fixed-effects model reduced the standard error of the
regression by more than half. Given the nature of the underlying model, a fixed-
effects model would be expected to be more appropriate than a random-effects
model. For some of the equations, this assumption was tested using the Hausman
test and the random-effects model was consistently rejected in favor of a fixed-
effects model.
The individual regulatory variables are sizable and with high t-values, though
the coefficient on civil service pay has the opposite sign to the one predicted
(table 2).Debt and industry control variables had consistently low t-values and
are dropped in subsequent regressions.
TABLE 2. Basic Static Generation Capacity Model Results
Explanatory Electricity Type of Staffing
Variables Law Regulator Funding Pay
Real DP per capita (log)
G
Electricity law
Autonomous regulator
License funding of
regulator
Civil service pay scales
nonmandatory
Debt payments
as a proportion of
national income
Industry value added as -0.001 -0.003 -0.002 -0.002
proportion of GDP (-0.249) (-0.920) (-0.645) (-0.599)
Estimation method Fixed effects Fixed effects Fixed effects Fixed effects
Adjusted R' 0.953 0.952 0.953 0.953
Standard error of 0.271 0.273 0.271 0.271
regression
F-statistic
Durbin-Watson
Number of observations
-- -
Dependent variable = log(e1ectricitygeneration capacity per capita).
Note: Numbers in parentheses are t-statistics.
Source: Authors' analysis based on data described in the text.
A sample average country is estimated to increase per capita generation
capacity in the long run by 18 percent through enacting an electricity law. In
this equation, as elsewhere, the long-run elasticity of per capita electricity gen-
eration capacity to per capita GDP is estimated at about 0.70-0.85. But the
equation clearly fits the data well and provides powerful initial support for the
importance of good regulation for generation investment.
The problem with the results in table 2 is that the high level of collinearity
between the regulatory variables implies that the coefficient estimates on the
individual effects are likely to be biased upward when taken in isolation. This
conjecture is confirmed when all four regulatory variables are included in a
single regression: the coefficient on the electricity law variable rises to 0.27,
and all the other variables become insignificant. Omitting the law variable led to
the funding variable becoming significant but with less than a 1 percent reduc-
tion in the standard error of the regression.
These results provide strong evidence that the high multicollinearity between
the regulatory variables significantly affects the coefficient estimates when
included in combination. The standard statistical solution to this problem is to
estimate a model using principal components-to better identify the effects of
the individual governance elements. The results showed that only the coefficient
estimate of the first principal component (accounting for 76 percent of the total
index variance) was statistically significant at the 5 percent level, with a t-value
of 3.8. The loadings of the individual components in the first principal compo-
nent were broadly similar to one another, and the loading on the electricity law
element was the highest.
The problemwith principalcomponents is that the resultsdo not necessarily have
any economic rationale. So, the preferred solution is to assemblethe four regulatory
variables into a regulatory index and to use that index as an explanatory variable.
This procedure, used in the Gutierrez (2003)telecommunicationsregulatory study,
has been used extensively in the literature on the economic impact of independent
central banks (see Geraats 2002 for a recent survey).
Thestandard procedure,adopted here, is to use asimpleadditiveindex. The index
takes the values O,1, ...,4 for each country in each year dependingon whether the
country scores1 or 0 on each of the four regulatory variables. However, as pointed
out by an anonymous referee, this procedureimposesthe restrictionthat each of the
variables included in the index has the same proportionate impact on the dependent
variable. This is a strong and highly debatable assumption, but at least our index is
derived from direct observation rather than from impressionistic indicators.14 In
view of this and other concerns, the summary results are reported estimating
alternative current and lagged versions of various regulatory variables.
The fixed-effects equations were all estimated with per capita GDP and the HZG
privatization variables as controls (table 3). In all the regressions, the coefficient
14. Estimation with a Guttman hierarchical index produced very similar results to those using a
simple additive index.
Cubbin and Stern 125
TABLE 3. Static Generation Capacity Model with Alternative
Regulatory Variables
Lagged 3-Year Lagged Quadratic
Explanatory Regulatory Regulatory Plus Electricity Electricity in Age of
Variables Index Index Regulator Law Law Regulator
Regulatory index(,) 0.022 (1.5)
Regulatory index(,_3) 0.041 (2.3)
Independent or 0.164 (2.9)
ministry regulator
in place 3 years
or more
Electricity act(,) 0.116 (2.6)
Electricity act(,-3) 0.143 (2.8)
Age of regulator 0.044 (3.6)
Age of regulator -0.0018(-2.8)
squared
Estimation method Fixed Fixed Fixed Fixed Fixed Fixed
effects effects effects effects effects effects
Adjusted R' 0.95 0.97 0.95 0.95 0.97 0.96
Standard error 0.27 0.21 0.27 0.27 0.20 0.21
of regression
Durbin-Watson 0.17 0.17 0.29 0.17 0.30 0.16
Number of observations 557 476 557 557 476 557
Dependent variable = log(e1ectricitygeneration capacity per capita).
Note: Numbers in parentheses are t-statistics.
Source: Authors' analysis based on data described in the text.
on per capita GDP was about 0.7, with a t-value of 8 or more. Here, attention is
concentrated on the results concerning the alternative measures of regulatory
governance.
Apart from the unlagged index, all the regulatory variables are positive and
significantlydifferent from 0 at the 5 percent level or better. The lagged variables
(including the 3-year plus dummy) are all larger than the contemporaneous
indicators and have higher t-values. The implication that it takes time to build
up the effect on regulation is supported in the age quadratic model, where the
maximum regulatory impact is estimated to be at about 14 years.
The dynamic modeling concentrates on the 3-year lagged index in column 2 of
table 3 and the 3-year plus regulator dummy in column 3. The lagged regulatory
law has the best overall fit, but given its collinearity with the other regulatory
variables, it is less satisfactoryoverall as a descriptor of the regulatoryframework.
The lagged index variable implies a maximum impact on per capita genera-
tion capacity of 16 percent-the same as for the 3-year regulator dummy. Note
that the latter includes ministry regulators as well as autonomous ones. But
particularly toward the end of the period, many ministry regulators were oper-
ating with powers and duties specified in a regulatory law.
One concern about the results in tables 2 and 3 is the low value of the Durbin-
Watson statistic. In the static form of the model, this would not be expected to
lead to biased coefficient estimates, but it may lead to overestimated t-statistics.
As a preliminary test, the column 3 model was estimated incorporating a first-
order autoregressive process. The coefficient on the lagged residuals was 0.79,
with a t-value of 42.8. The estimated coefficient on the lagged regulatory index
was both positive (0.02)and statistically significant at the 5 percent level (witha
t-value of 2.1) and was also positive and statistically significant for per capita
GDP. The estimated Durbin-Watson was 1.74.
The result suggests that autocorrelation in the static model does not signifi-
cantly affect either the coefficient estimates or their statistical significance. This is
explored more fully when explicitlyconsidering the results from dynamic models.
Econometric Results for Models of Generation Capacity: Static Model
Privatization and Competition
Incorporating the Domah data on privatization and competition into the model
did not produce any significant effects, but those data had major weaknesses.
This section reports estimates, using the better HZG data, of the relevant coeffi-
cients in fixed-effects regressions with the 3-year plus regulatory dummy vari-
able as the measure of regulation. In all cases, the estimated overall fit of the
equation and the coefficients on per capita GDP and the 3-year plus regulatory
dummy variable was within 1 percent of those in table 3. Results using the
regulatory index were similar.
On competition, the results reported above and others show consistent and
significant long-run effects on generation capacity levels of about 10-15 percent.
But the competition variable as defined provides no information on the amount
of private company electricity generated for sale-let alone whether it was from
an independent power producer selling to a single buyer or more from liberalized
wholesale markets. So, the results are more likely to indicate a degree of country
commitment to electricity reform rather than any genuine economic impact of
competition in generation markets.'' Sixty-one percent of the observations
scored 1 on this variable.
On privatization, the results varied. Unsurprisingly, there was no evidence
that minority privatization had any significant effect on generation capacity
levels. But neither did full privatization-though it applied to only 2.3 percent
of observations. There was some evidence, weak in statistical significance, that
majority privatization had a long-run positive effect on generation capacity
levels of about 8-10 percent.
In tests for interaction effects between the regulatory variables and both the
privatization and "competition" variables, none was significant at the 10 percent
level or better.
15. In many countries, including the U.K., the legal right for new entrants to generate for resale was
the first step in electricity reform-but achieved little or nothing in itself.
Cubbin and Stern 127
Econometric Results for Models of Generation Capacity:
Static Model-Country Governance Effects
On country governance, the Kaufmam indexes for rule of law and corruption by
country in 1998 were first included as explanatory variables. The corruption index
was never significant in the fixed-effectregressionsat the 5 percent level or better,
either as a separate variable or when interacting with regulatory variables. Esti-
mated coefficients on the Kaufmam rule of law index were ncver statistically
significant in their own right, but they sometimes approached significance when
interacting with the regulatory variables (Kaufrnann,Kraay, and Mastruzzi2005).
The Kaufmann rule of law index was, however, highly significant in an OLS
equation-leading to the nonsignificance of the electricity regulatory variable.
This last result (together with the relative constancy of the cross-country rank-
ings of general country governance indicators over long periods) is a major
reason why the estimated fixed effects may well capture a large part of the
countrywide institutional differences. On this last point:
No statistically significant correlation was found between the fixed effects
and the Kaufmann rule of law index.
But a sizable and statistically significant interaction term between the
regulatory index and the Kaufmann rule of law index was found in a
random-effects specification (a coefficient of 0.07, with a t-value of 2.3).
These results provide interesting pointers to the role of governance effects in
the model but are clearly far from conclusive.
A further test included in the static model the values of the World Bank
CHECKS index, a time-varying index of political risk. The index "counts the
number of veto players in a political system, adjusting for whether these veto
players are independent of each other, as determined by the level of electoral
competitiveness in a system, their respective party affiliations, and the electoral
rules" (Beck and others 2000, p. 28). The index yields a minimum score in the
absence of an effective legislature. The index score then increases linearly with
the addition of subsequent veto points.16 The index is available for all sample
countries for almost all years 1980-2000.
Including this index in the regressions rather than the single-year Kaufmann
index is a much stronger test of whether the estimates of our electricity regula-
tory governance effects are biased because of the absence of explicit country
governance measures. If country governance measures vary over time, as appears
to be the case, the impact is not captured either by the country-specific fixed
effects or by inclusion of the Kaufmann index for 1998.
Including the CHECKS index in the equations confirms the robustness of the
estimates reported in table 3. The estimated coefficient on the CHECKS index is
16. For further details on the definition of the index, see Beck and others (2000).We are grateful to an
anonymous referee not only for the suggestion that we include the index in our modeling but also for
providing the data for the countries and time periods in our sample.
correctly signed (positive)and about 0.015-0.020, with t-values of about 2.1 (an
increase of 1 point on the CHECKS index increases expected per capita genera-
tion capacity in the long run by 1.5-2.0 percent). Lagging the CHECKS index
variable has a very small impact on the value on the coefficients for itself or any
other variable.
The coefficient estimate on the (3-year lagged) electricity act was slightly
reduced by adding the CHECKS index (from 0.20 to 0.16), but its significance
remained high with an estimated t-value of 4.1. For the 3-year lagged regulatory
index, the coefficient on the regulatory variable was reduced from 0.058 to
0.049, but the t-value again remained high at 3.5. In addition, the coefficients
on privatization and "competition in generation" were virtually unchanged from
the results in table 4. This leads to the conclusion that the impact of the
electricity regulatory governance variables is genuine and not just a proxy for
variations in country governance.
Interaction effects between the CHECKS index and the electricity regulatory
governance variables were also tested for. The estimated coefficients were both
small and not significantly different from 0 (with t-values of about 0.6).
Last, an equation was estimated for per capita generation capacity including
the CHECKS index and per capita GDP but omitting any electricity regulatory
variable. The resulting coefficient estimate for the CHECKS variable increased
only slightly (to about 0.025), indicating that the collinearity between the
electricity regulatory variables and the CHECKS index was very small.
These results show that both sectoral regulatory governance and country
governance significantly affect the level of investment in per capita generation
TABLE 4. Static Generation Capacity Model: Alternative Privatization and
Competition Variables
All privatization Minority and
Explanatory and Competition Majority Privatization 50% or More
Variables Variables Variables Only Privatization
Minority
privatization
Majority
privatization
100% privatization
Majority or
full privatization
Competition (legal
right to generate
electricity for resale)
Dependent Variable = log(e1ectricitygeneration capacity per capita).
Note: Numbers in parentheses are t-statistics. Other independent variables in the regression are
per capita GDP and existence of 3-year plus regulator.
Source:Authors' analysis based on data described in the text.
Cubbin and Stern 129
capacity, but that the impact of the sectoral variables is several orders of
magnitude larger. Furthermore, the effects appear to be empirically separable
at least for electricity, an issue returned to in the conclusion.
Dynamic Models and Autocorrelation
Given the nature of the investment planning and construction, lags would
be expected to be quite long. Two main concerns were to establish whether
the results, particularly for the regulatory variables, were genuine causal
processes or merely spurious regression and to consider autocorrelation expli-
citly within a dynamic modeling framework rather than as a statistical auto-
correlation "corre~tion.~'To test whether the estimated long-run static
fixed-effects levels equations are genuine rather than spurious, check whether
there appears to be a plausible adjustment process. The levels equation can be
written as:
which can be estimated as:
where Yit= log(e1ectricitygeneration capacity per capita), Git= ~ O ~ ( G Dper
P capita),
Ritis a regulatorygovernance variable, and fi is the fixed effect for country i.
Equation (3) calculates the implied steady-state equilibrium, or long-term
value of Yit,which can be written as:
Now, postulate a partial adjustment error correction mechanism, under which
the actual value of capacity changes by a constant proportion of last year's
deviation from the long-term value:
where (Kt-, - Y,Z-,) is last year's deviation from equilibrium.
Equation (5)can be estimated by taking the residuals uit from equation (3)
and estimating:
An alternative procedure is to estimate directly a differenced version of the
long-run relationship, including country-specificfixed effects
130 T H E W O R L D B A N K E C O N O M I C REVIEW, VOL. LO, NO. I
Since the particular interest is in the size and significance of the regulatory
variable, R impose the estimate of from the long-term levels equation (3)and
estimate:
Equations (6)and (8)yield alternative estimates of A, the speed of adjustment,
which can be compared. In addition, alternative estimates of y, the impact of
regulation, come from the levels equation (2) and from the associated differ-
enced equation (8).
The validity of this procedure depends on the stationarity of the data genera-
tion process. The Pesharan-Shin W-statistic was used to test for stationarity. By
applying this test to the differenced equation (8),with the regulatory index as a
measure of Rit,the test clearly rejects the presence of a unit root in the residuals
with a t-value of -8.05.'~
The key results:
The estimates of A, the speed of adjustment, were low at 0.12 but very similar
between the levels and differenced equations, both with t-values of 8.9.
The estimates of the impact of regulation in the differenced equation (8)
were positive and significant, with t-values of 2.0 for the 3-year plus
regulator and 3.2 for the regulatory index.
The estimated long-run impact on per capita electricity generation capacity
in the differenced equation was 24 percent for the 3-year plus regulator and
almost 40 percent for the regulatory index.
There was no evidence of serial correlation in the differenced equations
(Durbin-Watson of 1.78). The overall fit of the differenced equations was
good, with an adjusted R' of about 0.15 and F-statistics of 4.5 or higher.18
Further estimated versions of these equations were produced using the 3-year
lagged regulatory index but also including the HZG privatization and competition
variables. These indicate a faster overall speed of adjustment, with an estimated
error correction term of 0.24-0.27. But this still imply more than 5 years before
half the regulatory effect on capacity appears (table 5).
The equations in table 5 demonstrate positive and statistically significant
coefficients in both level and differenced equations not just for the (3-year
lagged) regulatory index but also for majority privatization and competition.
17. Even in the corresponding levels equation, the Pesharan-Shin W-statistic does not appear to
suggest nonstationarity in the residuals, implying that our generation capacity variable, GDP, and our
regulatory variables are co-integrated. Very similar results were obtained on the unit root test with
alternative definitions of the regulatory variable.
18. See Cubbin and Stern (2004)for the full results.
Cubbin (2ndStern 131
TABLE 5. Generation Capacity-Error Correction Models
Dependent Variable[Log (Per Capita Generation Capacityi]
Explanatory variables Levels 1 Differences 2 Differences 2
Log real DP per capita
G
Index of regulatory governance
(0-4) (t-3)
Index of regulatory
governance (t-4)
Lagged residuals from 1
Error correction term
Majority privatization
Majority privatization (t-1)
Legal right of independent power
producer sales
Legal right of independent power
producer sales (t-1)
Estimation method Fixed effects Fixed effects Fixed effects
Adjusted R' 0.974 0.413 0.416
Standard error of 0.205 0.076 0.076
regression
F-statistic
Durbin-Watson
Number of observations
Note: Numbers in parentheses are t-statistics.
Source: Authors' analysis based on data described in the text.
The estimated long-run impact on per capita generation capacity from the
privatization and competition variables, derived from the differenced equation
in column 3, is 19 percent for the maximum regulatory index score, 20 percent
for a majority privatization, and 14 percent for the legal right for competition.
These results are very similar to those from the static models in tables 3 and 4.
They provide strong support for the hypothesis that the impact of regulation and
privatization on generation capacity in developing economies is positive and
sizable bur that it takes some years to build up.
IV. ENDOGENEITYD CAUSALITYAN IN GENERATION CAPA.CITYM ODELS
This section examines the endogeneity and causality issues tlnat arise in the
context of the model.19
19. We are grateful to Richard Gilbert and Jean-Michel Glachant for helpful discussions on these
issues.
Endogeneity
Much of the literature on regulatory effectivenessexpresses concerns about the
endogeneity of countries choosing to have an independent autonomous regula-
tory agency and of the quality of that agency's governance (see, for instance,
Fink, Mattoo, and Rathindran 2002; Gutierrez 2003). Countries with better
(unobservable)governance have better functioning regulatory agencies because,
for example, they have socioeconomic characteristics that better support the rule
of law, contracts, and commercialization. The problem is that it is very difficult
to find good instruments-instruments that are both correlated with the sus-
pected endogenous variable and uncorrelated with the error term-so that they
can be treated as exogenous. The alternative is to try to model explicitly the
decision to adopt regulatory reform. But this is difficult, and so far the results of
such modeling have been interesting but not very successful (see, for example,
Gual and Trillas 2004).
It is possible, however, to use a rank-based instrument to test for endogeneity
and to derive an instrumental variable (rv)estimator to control for it.'' With this
procedure, the coefficient on the residuals of the equation with the rank-based
index in the basicstatic equation for per capita generation capacity has a t-value of
1.7, implying marginal evidence of endogeneity of the Cubbin-Stern regulatory
index. But instrumenting the Cubbin-Stern index by using its predicted value in
place of the actual value produces virtually identical results-an estimated coeffi-
cient of 0.047 with a t-value of 4.3 in the instrumented case as opposed to an
estimate of 0.049 and a t-value of 4.0 in the noninstrumented case.
Like in Edwards and Waverman (2004) and Gutierrez (2003), the results
show some weak evidence of endogeneity of regulatory governance quality but
very little change in coefficient estimates after correcting for it.
Causality
Do the regulatory governance coefficientestimates reported here have any causal
interpretation?That the core results are maintained even with 3-year lags on the
regulatory index and with sophisticated dynamic modeling strongly suggests an
underlying causal relationship.
Even if they are not statistical artifacts arising from failures to address
dynamics or endogeneity adequately, they may still be merely descriptions of a
past set of events that cannot be applied to future electricity regulatory govern-
ance changes in sample countries-let alone to the introduction or development
of electricity regulation in other developing ec~nomies.'~
20. See Edwards and Waverman (2004),who follow Evans and Kessides (1993).For further discus-
sion of the procedure and its application in this context, see Cubbin and Stern (2004),pp. 35-36.
21. For the reasons stated in sections I and 11, we would not wish to claim that they are applicable to
countries with an excess supply of generation capacity at any time after 1980. This would exclude the
Central and East European countries, the Commonwealth of Independent States, and almost all Organi-
zation for Economic Co-operation and Development countries.
Cubbin and Stern 133
One reason this issue arises is that the regulatory literature derived from Levy
and Spiller (1994)emphasizes country-specific constitutional, legal, economic,
and political differences as crucial for the success or failure of utility regulation.
So, a highly reduced form that abstracts from all those issues may well fail to
reflect the local issues that seem to be so important in practice.
The answer lies in country-specific fixed effects. With 28 countries each
having up to 21 years of data, estimates can be obtained of the fixed effects
that should capture most if not all of the factors identified by L,evyand Spiller
and the subsequent literature. The estimated impact of, say, enacting a regula-
tory law with an autonomous regulator in Chile or Sudan (both (countriesin the
sample) will be very different. That impact is the combination of the predicted
effect of the relevant regulatory variables and each country's predicted fixed
effect. The fixed effect for Chile is strongly positive, relative to the sample
average, that for Sudan is strongly negative.
In other words, the coefficients reported here are "highest clommon factor"
estimates of the impact of regulatory governance indicators, where the fixed
effects not just control for but effectively"wash out" all the Levy and Spiller and
similar factors, including non-time-varying cross-country differences in country
governance. This means that the regulatory governance effects reported here are
not just average cross-country sample effects-they refer to a country with
average scores on country-specific fixed effects, including country-governance
fixed effects. Moreover, they are the impacts that one might expect, looking
forward, for a country:
With an average country-specificfixed effect.
Implementing an average-quality law.
Establishing an average-quality autonomous regulator-and so on.
For such a country, one might expect that implementing a best quality elec-
tricity regulator would increase per capita generation capacity in the long run by
about 15-25 percent. But countries that reduce their political risk scores also
increase their expected per capita generation capacity levels over and above the
impact of sectoral regulatory governance impacts (and vice versa).
Two policy implications emerge. First, the quality of overall country govern-
ance matters considerably for the impact of regulation on outcomes (as in the
rule of law).Second, countries cannot expect to achieve the gains estimated here
by enacting low-quality regulatory laws or introducing autonomous regulatory
agencies with very low staffing levels.22The corollary is that the potential gains
from introducing an electricity regulator could be significantly higher than the
average for countries with good overall governance that delliberately try to
introduce best-practice regulatory agencies and practices.
22. See Domah, Pollitt, and Stern (2002)for a full discussion of regulatory staffing issues and their
implications for the costs and effectiveness of regulation.
The higher potential gains from good regulatory institutions follow not just
from the logic of the fixed-effect modeling-they are confirmed by the signifi-
cance (and orthogonality) of the political risks index in our
The results of this study seem to provide a broadly consistent picture that a
regulatory agency with good governance characteristics can not only improve
regulatory outcomes in principle but that it seems actually to do so in practice.
For electricitysupply industries in 28 developing economiesover 1980-2001, an
index of regulatory governance is a consistently positive and statistically signifi-
cant determinant of per capita generation. The results, using fixed-effects esti-
mation methods, are similar to those found for telecommunications in
developing economies (Gutierrez 2003).
The main findingsfor per capita generation capacity in developingeconomies:
The effects of enacting a regulatory law, having an autonomous regulator,
and using license fees to fund the regulatory agency were each positive and
statistically significant at the 1 percent level.
Averaging over developing economies regulatory agencies, the estimated long-
run impact of the preferred measures of regulation is on the order of 15-25
percent,allother thingsequal,aftercontrollingforcountry-specificfixedeffects.
The effects on per capita generation capacity are robust to modeling with a
dynamic error correction model and to instrumental variable modeling to
allow for potential possible endogeneity biases.
On privatization and competition, there was some evidence of the effects on
generation capacity of majority privatization and of competition (itslegal intro-
duction). But the effects of competition are almost certainly more a reflectionof
a country's commitment to electricity reform than of a genuine market effect. A
positive and well-determined impact of majority privatization was found in the
dynamic modeling.
The strengthand robustness of the results were surprising. And since the regula-
tory changes typically took place in the mid- to late-1990s and the regulatory
variablesare mostsignificantwhen includedwith a 3-year lag, the resultsareclearly
not capturing just the Asian independent power producer boom or the Latin Amer-
ican privatization boom. But the recentness of the regulatorychanges may account
for thesurprisinglyhighsignificanceof theimpactof passingaregulatorylawrelative
to that of havingan autonomous regulator. It will be interestingto see whether the
estimatedeffect of havingan autonomous regulatoris higher in 5-10 years.
23. This view is also supported by the strong impact of the Kaufmann rule of law index variables in an
OLS equation and the highly significant, positive coefficient of an interactive governance-regulation
variable in a random-effects equation.
Cubbinand Stem 135
This article concentrated on the role of regulatory quality for capacity
and investment in the electricity industry. The results are very similar in
type to those previously found for telecommunications, and similar
approaches could likely be used to examine the institutional underpinnings
for investment in other infrastructure. The approach here might also be useful
for exploring the institutional contribution to capacity expansion for other
industries. That, however, would require long data sets to allow robust
estimates of country-specific fixed effects-and country governlance measures,
including the estimation of sophisticated dynamic models, such as error
correction models.
This article presented evidence suggesting that good regulatory governance
does have a positive and statistically significant effect on some electricity indus-
try outconles in developing economies-notably per capita generation capacity
levels. But it did not examine why this is so.
Examiningwhy and how regulationoperates to improve outcomesis not a task
that obviously recommends itself to econometric analysis. At this stage, that is
better pursued by case studies, with econometric work concentrated on whether
the results reported here are confirmed in subsequent analysis with superior data,
particularly for regulatory practice, privatization, and competition.
The results here are consistent with the literature on the role of institutions in
economic growth and with good country governance. Indeed, the evidence
suggests that good country governance and specific regulatory effectiveness are
mutually reinforcing. Both the quality of the electricity regulatory framework
and the quality of country governance (asmeasured by a political risk indicator)
are strongly associated with higher capacity, but as one might expect, the
sectoral variables have a markedly larger impact.
The key points are that regulatory agencies with better governance are:
Less likely to make mistakes.
More likely to correct mistakes speedily.
Less likely to repeat mistakes.
More likely to develop procedures and methodologies that involve partici-
pants and to develop good practice.
More likely to copy and implement best practice from other countries.
All these attributes reduce uncertainties for commercially operating compa-
nies, particularly private and foreign companies. This is especially important to
sustain and encourage long-lived, sunk investments in highly capital-intensive
industries. Regulatory agencies that maintain good governance thus provide an
effective underpinning for the operation of contracts and for the sound regula-
tion of monopoly elements.
These conclusions are also likely to be relevant for the institutional under-
pinnings of other sectors where raising sustained levels of long-lived or sunk
investments is important.
FIGURE A- 1. Continued
Ethiopia Grenada
? * r w Tsar
India I
& O.OOb00
1971 1480 1485 1940 1445 2bOO ZOOS
t..~
Mexico
Tsrr
Cubbin and Stern 139
TAB LE A- 1. Countries in Sample
Country Year of Regulatory Start Autclnomous Regulator
Argentina 1993 Yes
Barbados n.a. No
Bolivia 1995 Yes
Brazil 1997 Yes
Chile Pre-1980 N o
Colombia 1993 No
Costa Rica Pre-1980 Yes
Dominican Republic 1999 Yes
Ecuador 1997 Yes
El Salvador 1997 Yes
Ethiopia 2000 No
Grenada 1995 N o
India 1999 No
Indonesia n.a. No
Jamaica 1996 Yes
Kenya 2000 Yes
Malaysia 1991 N o
Mexico 1996 Yes
Nicaragua 1996 Yes
Nigeria 2001 N o
Paraguay n.a. No
Peru 1994 Yes
Philippines 1988 Yes
Sudan 2001 No
Trinidad Pre-1980 Yes
Uganda 2000 Yes
Uruguay 1998 Yes
Venezuela 1999 No
Note: ~ i G the time necessary to establish a functioning regulatbry entity, the year of the start
n
of regulation was typically taken as the year after the enactment of the relevant law.
Source: Authors' compilation based on data described in the text.
Beck, T., G. Clarke, A. Groff, P. Keeler, and P. Walsh. 2000. "New Tools and New Tests in Comparative
Political Economy: The Database of Political Institutions." Policy Research Working Paper 2283.
Washington, D.C.: World Bank.
Cubbin, J. S., and J. Stern. 2004. "Regulatory Effectiveness: The Impact of Good Regulatory Governance
on Electricity Industry Capacity and Efficiency in Developing Countries." Policy Research Working
Paper 3536. Washington, D.C.: World Bank.
Dahl, C. A., and C. Roman. 2004. "A Survey of Energy Demand Elasticities." IAEF, Working Paper.
Cleveland, OH: International Association of Energy Economics.
Domah, P., M. Pollitt, and J. Stern. 2002. "Modelling the Costs of Electricity Regulation: Evidence of
Human Resource Constraints in Developing Countries." London BusinessSchool Regulation Initiative
Working Paper 49 and CMI Working Paper 11. London, U.K.: London Business School, Regulation
Initiative and Cambridge, U.K.: Cambridge-MIT Institute, Electricity Project.
~dwards,G. E., and L. Waverman. 2004. "The Effectsof Public Ownership and Regulatory Independence
on Regulatory Outcomes: A Study of Interconnect Rates in EU Telecommunications." London Busi-
ness School EURI Working Paper. London, U.K.: London Business School, European Union Regula-
tory Institutions.
Evans, W., and I. Kessides. 1993. "Localised Market Power in the US Airline Industry." Review of
Ecotzomics and Statistics 75(1):66-73.
Fink, C., A. Mattoo, and R. Rathindran. 2002. "An Assessment of Telecommunications Reform in
Developing Countries." Working Paper 2909. Washington, D.C.: World Bank.
Geraats, P. M. 2002. "Central Bank Transparency." EconomicJozrral112(483):F532-65.
Gual, J., and F. Trillas. 2004. "Telecommunications Policies: Determinants and Impact." CEPR Discus-
sion Paper 4578. London, U.K.: Centre for Economic Policy Research.
Gutierrez, L. H. 2003. "The Effect of Endogenous Regulation on Telecommunications Expansion and
Efficiencyin Latin America." Journal of Regulatory Economics 23(3):257-86.
Henisz, W. J., B. A. Zellner, and M. F. Guillen. 2004. "International Coercion, Emulation and
Policy Diffusion: Market-Oriented Infrastructure Reforms, 1977-2000." William Davidson
Institute Working Paper 713. Ann Arbor, MI: University of Michigan, Stephen M. Ross Business
School.
Jamasb, T., R. Mota, D. Newbery, and M. Pollitt. 2005. "Electricity Reform in Developing Countries: A
Survey of Empirical Evidence on Determinants and Performance." Policy Research Working Paper
3549. Washington, D.C.: World Bank.
Kaufmann, D., A. Kraay, and M. Mastruzzi. 2005. "Governance Matters IV: Governance Indicators for
1996-2004." Policy Research Working Paper 3106. Washington, D.C.: World Bank.
Levy, B., and P. T. Spiller. 1994. "The Institutional Foundations of Regulatory Commitment: A Com-
parative Analysisof Telecommunications Regulation." Journal of Law, Economics, and Organization
10(2):201-46.
McMillan, M. 2001. "Why Kill the Golden Goose? A Political Economy Model of Export Taxation."
Review of Economics and Statistics 83(1):170-84.
Noll, R. G. 2000. "Telecommunications Reform in Developing Countries." In A. 0 . Krueger, ed.,
Economic Policy Reform: The Second Stage. Chicago: University of Chicago Press.
North, D. C. 1990. Institutiom, Institutional Change and Economics Performance. Cambridge, U.K.:
Cambridge University Press.
Pargal, S. 2003. "Regulation and Private Sector Investment in Infrastructure: Evidence from Latin
America." Policy Research Working Paper 3037. Washington, D.C.: World Bank.
Rodrik, D. 2003. "Growth Strategies." In P. Aghion, and S. Durlauf, eds., Handbook of Economic
Growth. Amsterdam: North-Holland Press.
Rodrik, D., A. Subramanian, and F. Trebbi. 2002. "Institutions Rule: The Primacy of Institutions Over
Geography and Integration in Economic Development." John F. Kennedy School of Government
Working Paper. Cambridge, MA: Harvard University.
Smith, W. 1997. "Utility Regulators-The Independence Debate." Public Policy for the Private Sector
Note 127. Washington, D.C.: World Bank.
Stern, J., and J. S. Cubbin. 2004. "Regulatory Effectiveness: The Impact of Regulation and Regulatory
Governance Arrangements on Electricity Outcomes-A Review Paper." Policy Research Working
Paper 3535. Washington, D.C.: World Bank.
Stern, J., and S. Holder. 1999. "Regulatory Governance: Criteria for Assessing the Performance of
Regulatory Systems; An Application to Infrastructure Industries in the Developing Countries of
Asia." Utilities Policy 8(1):33-50.
Wallsten, S. 2002. "Does Sequencing Matter? Regulation and Privatization in Telecommunication
Reforms." Policy Research Working Paper 2817. Washington, D.C.: World Bank.
Cubbin and Stern 141
World Bank. 1994. Infrastructure for Development. Washington, D.C.: World Bank.
-. 2004. World Development Indicators. Washington, D.C.: World Bank.
Zhang, Y. F., C. Kirkpatrick, and D. Parker. 2002. "Electricity Sector Reform in Developing
Countries: An Econometric Assessment of the Effects of Privatisation Competition and Regula-
tion." Manchester CRC Working I'aper 31. Manchester, U.K.: Centre for Regulation and
Competition.
An Em~iricalAnalvsis of State and Private-Sector
I d'
Provision of Water Services in Africa
Colin Kirkpatrick, David Parker, and Yin-Fang Zhang
Under pressure from donor agencies and international financial institutions such as the
World Bank, some developing countries have experimented with the privatization of
water services. This article reviews the econometric evidence on the effects of water
privatization in developing economies and presents new results using statistical data
envelopment analysis and stochastic cost frontier techniques and data from Africa. The
analysis fails to show evidenceof better performance by private utilities than by state-
owned utilities. Among the reasons why water privatization could prove problematic in
lower-income economies are the technology of water provision and the nature of the
product, transaction costs, and regulatory weaknesses.
The provision of safe and affordable water services is a priority for most
developing economies. According to the World Bank (2003, p. I),more than
1 billion people in the developing world lack access to clean water and nearly
1.2 billion lack access to adequate sanitation services. An estimated 12.2 million
people die each year of diseases directly related to drinking contaminated water.
The inclusion of a water access target in the Millennium Development Goals-
to halve the proportion of people without access to safe drinking water by
2015-is a recognition of the importance of safe water supply in reducing
poverty in the developing world (Calderon and Serven 2004).
A major cause of poor access to water services in developing countries is the
inefficiencies of water utilities, which serve mainly urban areas. In many sys-
tems, as much as a third of production is lost, revenues are insufficient to cover
operating costs, and the quality of the water is poor (World Bank 2004b,
p. 220). Faced with the deterioration in water sector performance, and with
Colin Kirkpatrick is the Hallsworth Professor of Development Economics at the Institute for Devel-
opment Policy and Management, School of Environment and Development, the University of Manchester,
UK; his email address is colin.kirkpatrick@man.ac.uk. David Parker (correspondingauthor) is Professor of
Business Economicsand Strategyat the School of Management, CranfieldUniversity, UK; his emailaddress
is david.parker@cranfield.ac.uk.Yin-Fang Zhang is a Lecturer in Economics at the Institute for Devel-
opment Policy and Management, School of Environment and Development,University of Manchester, UK;
her email address is yin-fang.zhang@man.ac.uk. The authors thank three anonymous referees and the
editor of the journal for their comments on earlier drafts.
THE WORLD BANK ECONOMIC REYIEW VOL. 20, NO. 1, pp. 143-163 doi:lO.l093/wber/lhjOOl
Advance Access publication January 19, 2006
O The Author 2006. Published by Oxford University Press on behalf of the International
Bank for Reconstruction and Development / THE WORLD ANK. All rights reserved. For permissions,
B
please e-mail: journals.permissions@oxfordjournals.org.
most water utilities under public ownership, donor agencies have advocated
privatization to promote more efficient operation, increase investment and
service coverage, and reduce the financial burden on government budgets
(World Bank 1995).
A range of services including water supply have now been opened up to
private capital (Harris 2003; World Bank 2003). Private participation has
been less common in water systems, however, than in other infrastructure
sectors, and the pace of reform has been slower and harder to sustain politically
(World Bank 2004b, p. 220). Although privatization appears to have the
potential to improve water services and meet the needs of the poor, these
goals may be difficult to achieve. The technology of water provision (high
fixed costs and location specificity)severely restricts prospects for competition,
the transaction costs of organizing long-term concession agreements are con-
siderable, and regulatory weaknesses suggest the need for caution. There is also
the difficultyof balancing adequate returns to investors and ensuring that water
services remain affordable to the poor.
The challenge for public policy is to meet both efficiency and social welfare
objectives and to determine whether or to what extent privatization is critical to
achieving the Millennium Development Goal for safe, accessible, and affordable
water services. This article explores these issues by examining the impact of
privatization of water services in Africa. It reviews the econometric evidence on
the impact of water privatization and then, for a data set for African water
utilities, uses statistical data envelope analysis and stochastic cost frontier
measures to triangulate the evidence and assess consistency across results.'
While data availability restricted the number of dimensions of performance
that could be estimated, the results for cost efficiency and service quality fail
to show that privatized water utilitiesperform better than state-run utilities.The
data deficiency may explain the failure to identify better performance under
private operation. However, special difficulties that face privatization and reg-
ulation in water services, also examined, likely play a role.
Private water suppliers have long been active as water vendors at the street level
in all developing countries, but there was little privatization of piped water
services before 1990 (Snell 1998; Collignon and VCzina 2000). Privatized ser-
vices could be found in only a few countries, generally French-speaking former
colonies such as CGte d'Ivoire that had inherited a reliance on private firms for
water services, as is the practice in France. Between 1984 and 1990, only eight
contracts for water and sewerage projects were awarded to the private sector
1. As Bauer etal. (1998) emphasize, there can be greater confidence in comparative analysis if
different measurements produce reasonably consistent conclusions.
Kirkpatrick, Parker, and Zhang 145
worldwide, and cumulative new capital expenditure in private water services
totaled less than $1 billion.
During the 1 9 9 0 however, there was a significant increase in water priva-
~ ~
tization, stimulated by donor agency pressures, and in 1997 private investment
had risen to $25 billion (World Bank 2003). By the end of 2000, at least
93 countries had privatizedsome of their piped water services, includingArgentina,
Chile, China, Colombia, the Philippines, South Africa, the transition economies
of Central Europe, and, among industrial countries, Australia and the UK
(Brubaker 2001). Based on the World Bank Private Participation in Infrastruc-
ture (PPI) Database for the period 1990-2002, there were 106 such projects in
Latin America and the Caribbean and 73 in East Asia and Pacific, but only
seven projects in the Middle East gnd North Africa, and 14 in Sub-Saharan
Africa. Latin America and the Caribbean and East Asia and Pacific together
accounted for more than 95 percent of total investment. During 1990-2002, a
small number of countries accounted for most of the privatization of water
services, and within these countries, the totals were dominated by a few large
contracts (table 1). In Argentina, one project accounted for $4.9 billion, or
20 percent, of all private investment in water services in Latin America, and in
the Philippines five contracts accounted for 38 percent of the investments in
East Asia.
Studies of privatization have found that competition is generally more impor-
tant than ownership itself in explaining improvements in performance in devel-
oping countries (Zhang, Kirkpatrick, and Parker 2003; Parker and Kirkpatrick
2005). But whereas competition is feasible in telecommunications and parts of
energy supply, such as generation, it is usually cost inefficient in the market for
water services. While there is scope for introducing some <:ompetition into
billing and metering and construction, replacement, and repair work within
water services, competition in the provision of water supplies is normally ruled
out by the scale of the investment in network assets that is needed to deliver the
TA BLE 1. Largest Investments in Water Services in Developin~gCountries,
1990-2002
Country Value (US$ Billions) Number of Projects
Argentina 7.23 10
Philippines 5.87 5
Chile 3.95 13
Brazil 3.17 33
Malaysia 2.75 6
China 1.93 44
Romania 1.04 3
Turkey 0.94 2
Indonesia 0.92 8
Source: Authors' analysis using data from the World Bank PPI Database (http://rrn.worldbank.
org/PPI).
146 T H E W O R L D BANK E C O N O M I C REVIEW, VOL. 20, N O . I
product. Moreover, even where competition for consumers might seem feasible,
for example, at the boundaries between different water utilities, the costs of
moving water down pipes is far higher than the costs of transmitting telephone
calls and distributing electricity, placing a serious limitation on competition.
Also, mixing water from different sources can affect water quality, an important
consideration for domestic consumers and especially for water-using industries,
such as brewing and food processing. In other words, the technology of water
supply and the nature of the product severely restrict prospects for competition
in the market and therefore the efficiencygains that can result from encouraging
competition following privatization. This leaves rivalry under privatization
taking the form mainly of competition for the market or competition to win
the contract or concession agreement.
Evidence suggests that privatization in noncompetitive markets produces
ambiguous results in terms of improving economic performance (Megginson
and Netter 2001), highlighting the need for effective regulation of privatized
utilities. The institutional requirements to ensure that privatized monopolies
perform well-an effective system of state regulation and supporting govern-
ance structures-are likely to be missing in many developing countries (Parker
and Kirkpatrick 2005). This represents a further difficulty to significantly
improving performance in the short term through water privatization.
Privatized water services contracts can be set up as service contracts for
specialized services (such as billing), management contracts and leases for
existing facilities (operatingexisting facilities without new private-sector invest-
ment), concessions (requiring private-sector investment in facilities), divestitures
(sale by the state of some or all of the equity in state enterprises), and greenfield
investments (including build-operate-transfer schemes) (Johnstone and Wood
2001; World Bank 2004b). The most common are contracts under which
private firms provide the services, but the government remains the ultimate
owner of the water system and may remain responsible for some investment
(OECD 2003). Of 233 water and sewerage contracts with the private sector
during 1990-2002 included in the World Bank's PPI Database, 40 percent
involved concession contracts, accounting for 64 percent of total investment.
Greenfield projects, less common, have often involved the building and opera-
tion of new water treatment plants, as in China, and build-operate-transfer
schemes for water supplies have been used in Latin America and the Caribbean.
Sales of state-owned water businesses to the private sector have been rare,
accounting for only 15.6 percent of water projects and 8 percent of the total
funds invested.
Although privatization of water services has occurred, it is important not to
exaggerate its importance. Little more than 5 percent of the world's popula-
tion receives drinking water through private operators (OECD 2003), and since
the Asian economic crisis of 1997198, there has been a marked slowdown in
infrastructure privatization in lower-income economies (Harris 2003). More-
over, the main forms that water privatization take raise concerns about the
Kirkpatrick, Parker, and Zhang 147
transfer of risk from the public to the private sector, an issue discussed later in
this article.
The case studyevidence on water privatization presents a mixeld picture, with
some cases showing improvements in labor ~roductivit~,operating costs, relia-
bility and quality of services, and share of the population served (Crampes and
Estache 1996; Estache, Gomez-Lobo, and Leipziger 2001; Galiani, Gertlier, and
Schargrodsky 2002; Shirley and Menard 2002; World Bank 2004b, pp. 252-57).
Balanced against these positive findings is some evidence of higher water charges
and public opposition leading to canceled schemes. The evidence is reviewed in
Kirkpatrick and Parker (2005)and by Shirley (2002).The few published econo-
metric analyses of the effects of water privatization in lower-income economies
present little evidence that privatization has resulted in marked improvement in
~erformance.Estache and Rossi (2002) compared private and public water
companies in 29 Asia and Pacific region countries using 1995 survey data on
50 water enterprises (22 with some form of private-sector participation) from
the Asian Development Bank. Adopting stochastic cost frontier modeling and
applying error components and technical efficiency effects models, they con-
clude that efficiency was not significantly different in the private and state water
sectors.
A study by Estache and Kouassi (2002),using a sample of 21 African water
utilities during 1995-97, estimated a production function from an unbalanced
panel data set and used Tobit modeling to relate resulting inefficiency scores to
governance and ownership variables. The study found that private ownership
was associated with a lower inefficiencyscore. However, only three firms in the
sample had any private capital, and levels of corruption and governance were
far more important in explaining efficiency differences between firms than was
the ownership variable.
A study of water supply in Africa in the mid- to late-1990s by Clarke and
Wallsten (2002)reported greater service coverage under private ownership. On
average, they found smaller supplies for lower-income households (proxied by
educational attainment) where there was a state-sector operator. While Clarke
and Wallsten conclude that private participation leads to more supplies to
poorer households, there may be offsetting service difficulties and higher
charges when supplies are privatized. Drawing strong conclusions on the desir-
ability of water privatization based on a single measure, such as service cover-
age, may be misleading. The analysis below uses a range of performance
measures in an attempt to address this problem.
To advance understanding of the results of privatization in wat~erservices, data
were taken from the Water Utility Partnership for Capacity Building in Africa's
Service Providers' Performance Indicators and Benchmarking Network Project
(SPBNE?) database, which includes 110 water utilities in Africa. The data
collected, usually by questionnaire survey, relate mainly to 2000.~The data set
used for this study covers 13 countries and 14 utilities that reported private-
sector inv~lvement.~However, not all of these firms could be included in each
stage of the analysis because of incomplete data entries. The descriptive statistics
for the sample are given in appendix table A.1.
Suppliers are categorized as either state owned or privately owned, a desig-
nation that captures the various institutional options for private-sector involve-
ment in the water sector, including management and leasing contracts. Ideally,
the form that private-sector involvement takes would be used to judge the
degree of privatization, but the data source permits ownership to be modeled
only as a binary variable. This limitation is shared by the earlier econometric
studies mentioned above. More generally, the data set is characterized by
heterogeneity, small sample size, and a small number of privatized firms. The
data limitations mean that the results must be treated as tentative.
Conclusions on the impact of water privatization may be sensitive to the
performance measure used. Therefore, to assess the impact of private capital on
performance in water services, a range of performance measures were calcu-
lated. First, several statistical measures were computed from the data set:
Labor productivity-ratio of labor costs to total costs, ratio of number of
staff to number of water connections, and staff per million cubic meters of
water distributed-to reflect the efficiency of labor use.
Proportion of operating costs spent on fuel and chemicals-to reflect
economies in nonlabor operating costs.
Rate of capital utilization-to reflect capital stock efficiency.
Average tariffs-to reflect the costs of services to consumers.
Share of the population served, unaccounted-for water (water losses), and
hours of availability of piped water per day-to reflect the quality of
service to consumers.
Average figures were computed for both state-owned and privately owned
water suppliers for between 61 and 84 utilities depending on the performance
measure (table 2). On average, the private sector seems superior in production
efficiency. Private-sector water utilities have higher labor productivity (lower
ratio of staff to number of connections and amount of water distributed) and a
lower share of labor costs in operating costs than do state-owned firms. The
2. The database (http://www.wupafrica.org) was developed with financial and technical support
from the UK Department for International Development. Data for a few utilities relate to 1999 or
2001. Given the closeness of the years, all data are treated as applying to 2000 to adopt a cross-sectional
analysis of performance.
3. Concession and management and lease contracts, together with privately owned assets, are
categorized as private utilities. The utilities classified as private were cross-checked with the World
Bank's PPI Database. The countries in the database with private water utilities are Cameroon, Cape
Verde, C6te d'Ivoire, Gabon, Ghana, Kenya, Morocco, Nigeria, Republic of Guinea, Senegal, South
Africa, Tunisia, and Zambia.
TAB LE 2. Performance Ratios in African Water Utilities, 2000
F-test for Between-Group
Average for State- Average for Private- Difference in means (Probability
Performance Indicator Sector Operations (SD) Sector Operations (SD) Statistics)
- - - -
Labor productivity
Labor costs in
total costs (percent)
Number of staff per 1,000 water connections
Number of staff per million cubic meters of water
distributed
Operating costs (percent)
+
e Share spent on fuel
Share spent on chemicals
Capital
Capital utilization (percent)
Consumer charges
Average tariff (US$ per cubic meter)
Share of customers metered (percent)
Quality of service
Share of population served (percent)
Unaccounted-for water (percent of total)
Avai!abi!ity of piped water (hours per day)
Source: Authors' analysis using data from the Water Utility Partnership for Capacity Building in Africa's Service Providers' Performance Indicators and
Benchmarking Network Project database (http://www.wupafrica.org).
150 THE WORLD BANK E C O N O M I C REVIEW, VOL. 20, N O . I
private sector is also more economic in its use of other inputs (fuel and chemi-
cals) and achieves a slightly higher capital utilization rate of 67 percent as
against 60 percent in the public sector firms.
Charges are on average 82 percent higher in the private sector, and more
customers have their water consumption metered under privatized service^.^ Me-
tering water can increase revenues derived from consumers by linking payments to
the volume of water used. The private sector also has lower water losses (probably
assisted by greater use of metering),averaging 29 percent as against 35 percent for
state-owned firms. Other measures of customer service suggest smaller differences
between the private and state sectors, however. On average, state-owned firms
supply piped water for 17 hours a day, while the private-sector records a slightly
lower figure of 16 hours. The state and private sectors serve about the same share
of the population in their areas, at 63 and 64 percent, respectively.
The standard deviations show a high degree of variance in performance
within both the state and the private-sector categories for each of the measures,
implying the need for care in interpreting conclusions based on average perfor-
mance. Similarly, the F-test results for the difference in means for the public and
private utilities' performance ratios show that none are statistically significant
(table 2). Also, data from the SPBNET database suggest that privately owned
water utilities in Africa are on average more than twice as large as state-
owned utilities in terms of the total volume of water distributed (92 million
cubic meters a day as against 36.4 million cubic meters a day) and have more
connections in their systems (averaging159,600 for private utilities and 94,500
for state-owned firms). This may ~artiallyaccount for the private utilities'
somewhat higher labor productivity.
To provide a fuller appraisal of relative performance, two further sets of
performance measures were calculated, drawing on the same database, one
using stochastic frontier analysis and one using data envelopment analysis.
Stochastic Cost Function Analysis
Because most water utility firms are required to meet demand and so are not
free to choose the level of output, the analysis is based on a cost frontier instead
of a production frontier. With output set exogenously, the firm is expected to
minimize the costs of producing a given level of output. The coefficients of the
cost function can be estimated by ordinary least squares (OLS) regression analy-
sis, or a stochastic cost frontier model can be estimated by the maximum
likelihood method. The stochastic cost frontier model decomposes the error
term into stochastic noise (vi)and cost inefficiency (pi).
Various distributions have been suggested for the inefficiency term in the
stochastic cost function. Two of the most commonly used are the half-normal
4. Tariff figures have to be viewed with care since tariff levels are affected by public policy toward
subsidies.
Kirkpatrick, lJarker, and Zhang 151
distribution (Aigner, Lovell, and Schmidt 1977) and the truncated normal
distribution (Stevenson1980). The truncated normal distributi~onis a general-
ization of the half-normal distribution, obtained by truncating the normal
e.
distribution at 0, with mean ,uand variance Preassigning p to be 0 reduces
the truncated distribution to half normal. The appropriate model for estimation
can be determined by testing the null hypothesis, Ho: ,u=0. If the hypothesis
p= O is rejected, the assumption of the truncated distribution is correct. If p is
not significantlydifferent from 0, a model assuming a half-normal distribution
should be estimated instead.
As in the parameterization proposed in Battese and Correa (1977), $and aZ,
are replaced by c? = o$ + 4,= $/(< + d),to
y allow application of max-
imum likelihood estimates. The parameter y lies between 0 and 1, with 0
indicating that the deviation from the frontier is due entirely to noise and 1
indicating that the deviation is due entirely to inefficiency. The superiority of a
stochastic frontier can be tested by the null hypothesis, Ho: y= 0. If the null
hypothesis cannot be rejected, this indicates that the inefficiency term should be
removed from the model, leaving a specification with parameters that can be
consistently estimated using OLS.
The stochastic cost function has been widely specified as a Cobb-Douglas
function or as a translog cost function. A generalized likelihood ratio test is used
to determine whether a Cobb-Douglas function is appropriate. The result shows
that the null hypothesis of the Cobb-Douglas specificationcann~otbe rejected. In
addition, Leamer's extreme bound analysisshows that the range of the coefficients
of the key variables for the Cobb-Douglas function is much smaller than that of
the translog mode, confirming that use of the Cobb-Douglas specification is
appropriate.s To account for variable returns to scale, the quadratic term of the
output variable is included. The coefficientis statistically insignificant, h~wever.~
A likelihood ratio test also points to the standard Cobb-Douglas specification.
As in the literature, the cost function is estimated using data on the cost level,
the output level, and input prices. Operating and maintenance costs (COST) are
used as the dependent variable in the cost frontier because adequate capital cost
data are not available to compute total costs. An arbitrary cost function is
therefore formulated that excludes the price of the capital input.7 Average
5. Leamer's extreme bound analysis was applied to the Cobb-Douglas and ithe translog specifica-
tions. Accor&ngly, the output and input variables were treated as focus variables and the control
variables as doubtful variables. The bounds from the Cobb-Douglas model were much narrower than
those from the translog model. In addition, in the translog model, the bound for the material input
variable spanned zero. The results suggested that the coefficientsfor the Cobb-Douglas model were more
robust than those for the translog specification.
6. The full results of these tests are available from the authors.
7. Estache and Rossi (2002)follow a similar procedure. In response to a referee's comment that the
exclusion of a fixed-capital measure might result in a misspecification of the cost function, an alternative
specification of the cost function was tested that used the number of water treatment plants as a proxy
variable for capital costs. The results for the ownership variable were unaffected.
personnel cost per employee (MP) is used to reflect the cost of labor, and material
cost per unit of water distributed (MAT) is included as an additional determinant
of noncapital costs. The amount of water distributed per year (WD) is included in
the cost function as the output variable. Also included is a quality variable,
measured by the hours of piped water available per day ( Q u ~ ~ ~ ~ ) . 8
A number of control or environmental variables are also included to capture
cross-country heterogeneity in the political, legal, and economic en~ironment.~
Good governance, in the form of sound finance and regulatory systems and
protection of property rights, has been found to be an important explanation for
differences in economic performance (North 1990; Jalilian, Kirkpatrick, and
Parker 2002; Kaufmann, Kray, and Zoido-Lobat6n 2002), including in water
services (Estache and Kouassi 2002). The freedom variable (FRD) developed by
the Fraser Institute (http://www.freetheworld.com) is therefore included to cap-
ture wider governance or regulatory effects on performance in water utilities
that might otherwise be attributed to ownership. An index of property rights
(PROPERTY) is used as a measure of the quality of the investment environment
(http://www.freetheworld.com). The fiscal balance variable (BALANCE) proxies
the quality of macroeconomic management (http://www.freetheworld.com). A
density variable, measured by population served per connection (DEN), drawn
from the SPBNET database, is included because it plays an important role in
defining the network infrastructure.1� Annual water resources per capita (WRS)
is used as another control (WRI 2003). DP per capita (GDP) is included to capture
G
the extent of economic development (World Bank 2002). Finally, a dummy
variable (ONS) is included to account for the effects of ownership on perfor-
mance, taking a value of 1 if the utility had private capital.
All variables except the ownership variable and those in index or percentage
terms are logged. In total, the estimations include 76 observations, including 10
private-sector operations. The program FRONTIER 4.1 is used to obtain the max-
imum likelihood estimates of the parameters and efficiency measures. The proce-
dure for estimation is as follows. An error-component model is first estimated
with the assumption of a half-normal distribution for the inefficiency term.'' To
8. Alternative quality indicators (unaccounted-for water and share of samples that fail to meet quality
standards) were also tested, with similar results.
9. See Rodrik, Subramanian, and Trebbi (2004)and Glaeser and others (2004)for a discussion of the
use of institutional quality variables in quantitative analysis.
10. As pointed out by a referee, this density measure does not fully capture the dispersion of
connections since it does not allow for the number of connections per building. Data on more common
measures of dispersion, such as connections per kilometer of main lines or connections per square
kilometer, were not available.
11. The error component model is the standard form of stochastic frontier model used in the
literature. It decomposes the error term into stochastic noise and cost inefficiency. The truncated-
distribution assumption yields p=0.47, with a standard error of 2.56. A likelihood ratio test shows
that the hypothesis p =O could not be rejected at the 10 percent level. Consequently, the results from the
model with the half-normal assumption were adopted.
Kirkpatrick, Parker, and Zhang 153
TABLE 3. The Stochastic Cost Frontier Results
Error-Component Model Technical Efficiency
(Half-Normal Distribution) Effects Model
Ordinary Maximum Ordinary Maximum
Variable Least Squares Likelihood Least Squares Likelihood
Constant 4.17 (2.60)""" 1.18 ( 1 . 6 ) 4.05 (2.47)"" 1.55 (0.29)
InWD 0.76 (13.22)"r" 0.88 (29.49)""" 0.76 (13.02)"'" 0.86 (23.97)'""
I~QUALI 0.12 (0.81) 0.14 (1.88)"' 0.06 (0.38) 0.11 (1.80)""
I m p 0.26 (3.76)"" 0.15 (4.33)"" 0.25 (3.62)+"* 0.17 (5.28)'""
I ~ M A T 0.56 (8.20)"'" 0.65 (15.84)""' 0.56 (7.99)""' 0.63 (8.25)"""
l n w ~ s -0.001 (0.01) -0.09 (1.48)" 0.0009 (0.01) -0.08 (0.22)
DEN -0.02(0.44) 0.00003 (0.001) -0.028 (0.65) -0.02 (0.10)
~ ~ G D P 0.09 (0.85) -0.01(0.26) 0.15 (1.40)' -0.03 (0.31)
FRD -0.13 (1.28) -0.08 (0.22) -0.12 (1.18) -0.02(0.54)
PROPERTY -011(1.38) -0.05 (4.03)'"'" -0.13 (1.59)" -0.06(1.83)":'
BALANCE 0.02 (0.64) -0.004(0.32) 0.02 (0.53) 0.004 (0.09)
ONS 0.42 (2.00)"* 0.15 (1.05) 0.11 (0.15)
JONS 0.11 (.015)
Y 0.98 (0.63E+07) 0.98 (0.21E+06)
Generalized 34.63 44.53
likelihood ratio test
Number of 76 76 76 76
observations
"Significant at the 10 percent level.
""Significant at the 5 percent level.
"""Significant at the 1percent level.
Source: Authors' analysis based on data described in text.
test the robustness of the results on ownership, a technical efficiency effects
frontier is then estimated in which the inefficiency effects are expressed as a
function of the ownership dummy variable.12
The value of y in the error-component model suggests a high ratio of the
variance of inefficiency to the total residual variance (0.98; table 3).13 Analo-
gously, the high value of y means that the stochastic frontier is superior to OLS
modeling in explaining the cost structure of water utilities (both results are
presented in table 3, for comparison). This is also confirmed by the generalized
likelihood ratio statistic, which exceeds the critical value at the 1percent level.14
The results of the half-normal error-component model show that the output
variable, water distributed annually (lnw~),has a positive and significant effect
12. The technical efficiency effects model can be used to investigate the determinants of technical
inefficiencies among firms. The technical efficiency effects frontier is a stochastic frontier model that
explicitly formulates technical inefficiency effects in terms of firm-specific factors. All parameters are
estimated in a single-stagemaximum likelihood procedure.
13. A referee has pointed out that the error term may be capturing more than just inefficiencywhere
there is misspecificationbecause of heterogeneity or measurement problems.
14. The critical value was obtained from Kodde and Palm (1986).
on operating costs. This is in line with expectations. Similarly, the variables of
service quality (~QUALI),labor price (Imp), and material cost ( I ~ A Tare all )
significant and correctly signed. The negative and statistically significant (at the
10 percent level)coefficientfor the water resource variable (lnw~s)is also consis-
tent with expectations. The costs of water production and distribution would be
expected to be lower in countries where water resources are abundant. The
negative coefficients of income per capita (I~GDP) and the freedom index (FRD)
suggest that the operational costs of the utilities may be lower in countries that are
wealthier, with sounder institutional governance. However, the effects are not
statistically significant. More robust evidence of the influence of institutional
development is provided by the property rights variable (PROPERTY), which shows
negative and significant effects on the cost level, indicating that costs are lower in
countries where property rights and therefore private investment are better pro-
tected. The impact of the government fiscal management measure (BALANCE)
appears to be trivial. Contrary to expectations, however, the resultsfor the density
variable ( ~ D E N )are statistically insignificant. The coefficient of the ownership
dummy variable (ONS) is positive, suggesting that private ownership is associated
with higher costs. However, the result is not statistically significant.
To assess the robustness of these results, a technical efficiency effects model is
estimated in which the inefficiency term is expressed as a function of the own-
ership dummy variable. In this model, the inefficiencyerror, pi, has a mean of mi
and mi=Sxi,where xiis a vector of variables that may influence the efficiency of
a firm. This is taken as the ownership dummy variable in the estimation. The
maximum likelihood estimation shows that the coefficient SONSis positive but
not statistically significant (table 3). This finding is consistent with the owner-
ship outcome from the error-component model.
The Data Envelopment Analysis
A data envelopment analysis was also undertaken.15Water distributed is repre-
sented by the volume of output produced, and the number of hours of piped water
available per day is used as the proxy for the quality of water services. (Unac-
counted-for water and the share of samplesthat failed to meet the quality standards
werealso usedasa proxyfor quality of service,and the resultswereverysimilar.)An
input-orientedvariablereturnstoscalemodel wasadopted to allowfor variations in
the sizeof the utilities.16The analysisincludes66 utilities, nine of them private. The
inputs are personnel cost per employee (becausenumber of staff would not reflect
the average skill level of staff17),material cost per unit of water distributed, and
number of water treatment works. The efficiency scores from the initial data
envelopment analysis are regressed on the control variables (DEN, WRS, GDP, FRD,
15. The authors are grateful to Catarina Figueira for assistance with the data envelopment analysis.
16. A constant returns to scale model produced a similar set of results but with lower overall scores.
17. The authors thank a referee for drawing this to their attention.
Kirkpatrick, Parker, and Zhang 155
TA BLE 4. Summary of the Data Envelopment Analysis Results
Utilities With TJtilities With
100 Percent Utilities With Utilities With Utilities With Less Than
Relative 90-99 Percent 80-89 Percent 70-79 Percent 70 Percent
Ownership Efficiency Efficiency Efficiency Efficiency Efficiency
State 32 (53) 7 (121 9 (36) 5 (9) 4 (5)
Private 6 (67) 1(11) 1 (11) 1 (110) 0 (0)
Values are expressed as n (%).
Note: The lowest score, 52.5, was recorded by a state-owned water utility in South Africa.
Source: Authors' analysis based on data described in text.
PROPERTY, and BALANCE as defined earlier) using a Tobit model. Only population
served by connection, DEN, and the property rights variable, PROPERTY, are statisti-
cally significant, and these two variables are included as control variables in a
second-stagedata envelopment analysis.ls
The final data envelopment analysis results were tabulated by efficiency
scores: the number of private and state utilities that achieved a score of
100 percent efficiency, 90-99 percent, and 80-89 percent (table 4).19 Signifi-
cantly, state-owned firms help to form the efficiency frontier, suggesting that
state ownership does not necessarily lead to low relative efficiency. More than
half of the state-owned firms in the data set (32of 57)were on the frontier. Six
of the nine private operations included in the analysis populated the frontier.
Therefore, the data envelopment analysis results appear to be consistent with
the stochastic frontier analysis in suggesting that the efficiency performance of
state-owned water firms in Africa is comparable to that of private enterprises.
However, the results provide stronger evidence for possible higher relative
efficiency in the private sector as a whole. For example, no utilities with
private-sector involvement have less than 70 percent relative efficiency, and
67percent of private as against 53percent of state operationspopulatethe frontier.
It should be noted, however, that there are only nine private firms in the sample.
With the results of the analysis presented here, it is interesting to consider why
privatization of water services may be problematic in lower-income economies.
The answer seems to lie in a combination of the technology of water provision
18. The inclusion of control variables in data envelopment analysis is widely practiced in empirical
studies; see, for example, Rugggiero (1996,2004j and Paradi and Schaffnit (2004).Wang and Schmidt
(2002),however, are critical of this two-step procedure in data envelopment analysis.
19. Data envelopment analysis provides scores relative to peers with similar operating characteristics
based on an estimated efficiency frontier. The resulting scores are relative, not absolute, scores. There-
fore, a score of 100 percent does not imply absolute efficiency but merely efficiency compared with the
other units in the analysis. Similarly, a stochastic cost frontier approach creates a frontier based on actual
performances in the data set.
and the nature of the product, the costs of organizing long-term concession
agreements, and regulatory weaknesses.
As explained, the technology of water supply and the nature of the product
severely restrict prospects for competition in the market and therefore the
efficiency gains that can result from encouraging competition following priva-
tization. This leaves rivalry under privatization mainly in the form of competi-
tion for the market-competition to win the contract or concession agreement.
However, serious problems can arise related to pervasive transaction costs in
contracting for water services provision. These include the costs of organizing
the bidding process, monitoring contract performance, and enforcing contract
terms where failures are suspected (Williamson 1985). The economics literature
suggests that such costs are likely to be high where there are serious information
asymmetries at the time of contract negotiation.
Information imperfections are especially likely when contracts have to be
negotiated to cover service provisionover long periods of time. Manyfuture events
that could affect the economic viabilityof the contract and the acceptability of the
service offering are unforeseen, and some may be unforeseeable. Concession
agreements in water are typically negotiated for 10-20 years or more. Inevitably,
therefore, thecontracts will need to permit periodicadjustment of such variablesas
price, volume, and quality during the contract life. The contract will be incomplete
in terms of specifying all of the contingencies that may trigger such adjustments
and the form the renegotiation might take. This requires considerable skills on the
part of both government and companies when operating water concessions, to
ensure that the outcome is as mutually beneficial as possible.
The usual approach in water concessions is a two-part bidding process:
selection of approved bidders based on technical capacity and then selection
of a winner based on such criteria as the price offered and the service targets.
However, the smaller the number of bidders, the greater the scope for actual or
tacit collusion in bidding and the less competitive will be the bidding process.
The evidence suggests that water concessions in developing countries are subject
to small-numbers bidding (McIntosh 2003, p. 2). For example, in 2001 in
Nepal, 18 companies expressed interest in operating a water contract in the
first stage of the process, but only two serious bidders remained in the final stage
(cited in Mitlin 2002, p. 17). In Argentina, there have usually been only a small
handful of applicants for water concessions, typically between two and four
(Estache2002).To stimulate greater interest, concessions can include sovereign
(government or donor agency) guarantees of profitability, but this introduces
obvious moral hazard risks-with profits guaranteed, what incentive does the
concession winner have to operate efficiently?
The literature on transaction costs also suggests that small-numbers bidding
is a source of opportunistic behavior (Williamson 1985),leading to both adverse
selection and moral hazard. Adverse selection takes the form of suboptimal
contracts at the outset, as one of the contracting parties acts opportunistically to
arrange especially favorable terms. Moral hazard occurs when one of the
Kirkpatrick, Parker, and Zhang 157
contracting parties renegotiates the terms of the contract in its favor during the
lifetime of the contract. During contract renegotiation, either the company or
the government could be the loser, depending on the results.20
Guasch (2004)concludes that 75 percent of water and sanitation concession
contracts in Latin America and the Caribbean were renegotiated significantly
within a few years of being signed. In Buenos Aires, prices were raised within
months of the start of the water concession (Alcazar,Abdala, and Shirley 2000).
But even the ability to renegotiateterms may not be sufficientto overcomeinvestor
reluctance to participate in water privatizations, thus reinforcing the small-
numbers bargaining problem. Difficulties arise especially when private investors
fearthat there is no long-term political commitmentto water privatization(Rivera
1996). Moreover, cronyism and corrupt payments to win concessions may com-
promise the legitimacy of the privatization process. For example, in Lesotho, the
Highlands Water Project was associatedwith allegationsof bribes to government
officials (Bayliss2000, p. 14).Esguerra (2002, p. 2)shows how t~hewater conces-
sions in Manila were backed by the Philippines' two wealthiest families with
support from multinationals: "It appears that the two companies' approach was
to win the bid at all costs, and then deal with the problemsof prlofitability later."
Studyingconcessioncontracts in developingcountries,Harris and others (2003)
find that water and sewerage concessions have the second highest incidence of
contract cancellation after toll roads. This is not surprising given the substantial
potential sunk costs in the water industry. Tamayo and others (1999, p. 91) note
that the specificity of assets in the water industry is three to four times that in
telecommunicationsand electricity. Handley (1997)stresses the problems caused
by inadequate risk management techniques in developing countries. The prefer-
ence by the private sector for the state to remain responsible for the infrastructure
in water contracting reflects the desire of companiesto minimizetheir sunk costs.
Transaction costs in water concessions reinforce serious weaknesses in govern-
ment-regulatory capacity in developing countries (Spillerand Savedoff 1999, pp.
1-2). For example, in India, there have been some local moves to attract private
capital into water supply, notably in Gujarat, Maharashtra, and Tiruppur. But
regulatorysystems are underdeveloped, and in Tiuppur, they appear to be largely
under the indirect control of the water operator (TERI 2003, pp. 171-221). As
Mitlin (2002, pp. 54-55)concludes on the experience in Manila:
The gains [from privatization] may be less than anticipated because the assumption that the
involvement of the private sector would remove political interference from the water sector was
wrong. It may be that processes and outcomes have simply become more complex because the
water supply industry now has the interests of private capital in addition to a remaining level of
politicisation and an acute level of need amongst the poorest citizens.
20. For example, in the concession involving Maynilad in Manila, the company terminated the
concession when it was refused a rate adjustment to which it felt entitled. By contrast, in Dolphin Bay,
South Africa, the municipality believed that it had little alternative but to agree to an unplanned price rise
when the private-sector supplier threatened to withdraw services (Bayliss2002, p. 16).
158 THE W O R L D BANK E C O N O M I C REVIEW, VOL. LO, NO. I
To assess the effects of regulation on water privatization in Africa, the
stochastic cost function analysis was repeated, this time incorporating a regu-
latory variable as a dummy variable alongside the freedom variable (represent-
ing wider good governance in a country). The SPBNET database provides
information on regulation of prices, water quality, and customer services. The
different regulatory indicators are included separately in the regressions and are
also combined into a composite regulation dummy variable to reflect the pre-
sence or absence of regulation in the water sector.
Regulation is expected to influence costs depending on its form. For example, a
good regulatory regime should create more investor certainty and may reduce the
costs of capital. Alternatively,regulation could raise costs by imposing higher and
more expensive quality standards or by raising uncertainty for investors. The
regression results show a negative sign for the composite regulation dummy
variable and for the water quality and services dummy variables, suggesting that
regulation lowered operating costs. However, these results are not statistically
significant.The regulation dummy variablefor tariff regulation is positivelysigned
and statistically significant, suggesting that regulation of prices increased costs.
The findings from this stage of the analysis are therefore inconcl~sive.~~Re-
gulation, both sector specific and as reflected in the general standards of
governance in a country, are statistically insignificant. The single exception is
related to tariff regulation, and the result is consistent with recent concerns that
state regulation can raise costs (World Bank 2004a). However, the regulation
variables used are far from ideal, and future research would benefit from
developing a set of superior regulatory variables that more closely reflect the
form of regulation rather than simply its existence.
IV. CONCLUSIONS
In principle, privatization has the potential to improve water services in devel-
oping countries, reversing decades of underinvestment and low productivity
- -
under state supply. However, the results, taken together, do not provide strong
evidence of differences in the performance of state-owned water utilities and
water utilities involving some private capital in Africa. While the data envelop-
ment analysis results point tentatively to private-sector superiority, the stochas-
tic cost frontier analysis provides some evidence that state-owned utilities have
better cost performance, but the results are statistically insignificant. The
descriptive statistics suggest no statistically significant differences.
The results therefore complement those of Estache and Rossi (2002), who
also failed to find evidence that the performance of privately owned water
21. The detailed results can be obtained from the authors. A Tobit model was used to assess the
impact of the regulation variables on the data envelopment analysis scores discussed earlier. The results
were also statistically insignificant.
Kirkpatrick, Parker, and Zhang 159
utilities in developingcountries is superior to that of state-owned firms. Estache
and Kouassi (2002) report a statistically significant result for the effect of
privatization. However, this is based on data for only three privatized utilities
in a total sample of 21 water utilitiesin Africa, and governance and institutional
factors were found to be much more significant in explaining performance.
Admittedly, the results here contrast with the findings of Crarnpes and Estache
(1996) and Galiani, Gertlier, and Schargrodsky (2002), who concentrate on
service coverage. They concludethat privatizationincreased the number of people
provided with safe water and sanitation. This study found no real difference in
the share of the population served between private and state-owned utilities, but
the limited availability of data precluded detailed exploration of this dimension of
service. As with any study, the findings are dependent on the data used, and these
were far from ideal. There is also the possibilitythat governments in Africa turned
to private capital for the worst performing water utilities, thus making it less
likely that the private sector would exhibit superior performance.
Other reasons why water privatization might prove problematic in lower-
income economies were also identified and may help to explain why this and
earlier studies have not found an unequivocally positive effect of private own-
ership on performance. Regulation dummy variables were included in the
stochastic cost frontier model to shed further light on the import:ance of regula-
tion, but most results were statistically insignificant. This outcome may reflect
the crudity of the regulatory variables used, which simply measure the existence
of water regulation not its impact on the management of utilities. Under condi-
tions of perfect competition, perfect information, and complete contracts, own-
ership does not matter (Shapiro and Willig 1990) and the regulatory
environment becomes inconsequential. However, none of these conditions
applies to water services, and governance and regulatory variables are expected
to be important in determining performance before and after privatization.
Finally, it needs to be stressed that providing affordable, safe, and accessible
water to the poor is a fundamental priority for low-income countries. Policy-
makers and regulators are likely to face difficult tradeoffs between ensuring that
poor households are provided with affordable water suppliesand allowing firms
to charge prices high enough to recover costs and attract the foreign capital and
technical capabilities needed to upgrade and expand the water supply network.
This study found that private operation of water facilities is associated with
much higher average water charges and with greater use of water metering. But
what are the impacts of this on water consumption and health? Water privati-
zation usually means the involvement of a handful of major international
companies. But what effect does this have on the development of indigenous
ownership and regulation of socially important assets?22Also, if privatization of
22. Kirkpatrick and Parker (2005)discuss the implications of liberalization of water senices under the
WorldTrade Organization'sGeneral Agreementon Trade inServicesfor domestic regulation of water utilities.
APP ENDIX
TA BLE A.1. Descriptive Statistics
Standard
Statistic Minimum Maximum Mean Deviation
Availability of piped 2.00 24.00 17.17 6.99
water (hours per day)
Labor cost per employee 134.49 88,478.92 12,806.64 17,851.00
(in PPP units)
Material cost per unit 0.00024 0.67 0.17 0.15
water distributed
(in PPP units)
Number of connections 0.01 526.14 61.78 100.34
(in thousands)
Total operating cost 62,812.45 1,107,688,842.80 53,038,864.01 157,294,171.22
(in PPP units)
Total volume of water 8200 668,000,000 48,258,663.55 95,605,864.54
distributed per year
(cubic meters)
Note: PPP is purchasing power parity.
Source: Authors' analysis using data from the Water Utility Partnership for Capacity Building in
Africa's Service Providers' Performance Indicators and Benchmarking Network Project database
(http:Nwww.wupafrica.org).
water servicesleads to full cost recovery, is this outcome compatible with poverty
reduction and what are the environmental implications of privatization? Clearly,
water privatization raises a complex set of political economy questions that
deserve fuller exploration than has been possible here because of data limitations.
Aigner, D., C. Lovell, and P. Schmidt. 1977. "Formulation and Estimation of Stochastic Frontier
Production Function Models."Journal of Econometrics 6(1):21-37.
Alcazar, L., M. Abdala, and M. Shirley. 2000. "The Buenos Ares Water Concession." Policy Research
Working Paper 2311. World Bank, Washington, D.C.
Battese, G. E., and G. S. Correa. 1977. "Estimation of a Production Frontier Model with Application to
the Pastoral Zone of Eastern Australia." AustralianJournal of Agricultural Economics 21(3):169-79.
Bauer, P., A. Berger, G. Ferrier, and D. Humphrey. 1998. "Consistency Conditions for Regulatory
Analysis of Financial Institutions: a Comparison of Frontier Efficiency Methods." Journal of Eco-
nomics and Business 50(2):85-114.
Bayliss, K. 2000. "The World Bank and Privatisation: A Flawed Development Tool." University of
Greenwich, Public Services International Research Unit, London.
. 2002. "Privatisationand Poverty: The Distributional Impact of UtilityPrivatisation."Centre on
Regulation and Competition Paper 16. University of Manchester, Centre on Regulation and Compe-
tition, Manchester.
Brubaker, E. 2001. The Promise of Privatization. Toronto: Energy Probe Research Foundation.
Calderon, C., and L. Semen. 2004. "The Effects of Infrastructure Development on Growth and Income
Distribution." Policy Research Working Paper 3400. World Bank, Washington, D.C.
Kirkpatrick, Parker, and Zhang 161
Clarke, R. G., andS. J. Wallsten.2002. "Universal(1yBad)Service: ProvidingInfrastructure Servicesto Rural
and Poor Urban Consumers."Policy Research Working Paper 2868. World Bank, Washington, D.C.
Collignon, B., and M. Vkzina. 2000. "Independent Water and Sanitation Providers in African Cities: Full
Report of a Ten-Country Study." United Nations Development Programme-World Bank Water and
Sanitation Program, Washington, D.C.
Crampes, C., and A. Estache 1996. "Regulating Water Concessions: Lessons from the Buenos Aires
Concession." Private Sector Note 91. World Bank, Washington, D.C. (http://wvvw.worldbank.orgl
DocumentsffublicPolicyJournal).
Esguerra, J. 2002. The Corporate Muddle of Manila's Water Concessions: How the World's Biggest and
Most Sticcessful Privatisation Turned Into a Failure. London: Water Aid.
Estache, A. 2002. "Argentina's 1990 UtilitiesPrivatization: A Cure or a Disease?"Paper presented at the
Conference on Applied Infrastructure Research. Berlin University of Technology, October 12, Berlin
(http://wip.tu-berlin.de.workshop).
Estache, A., and E. Kouassi. 2002. "Sector Organization, Governance, and the Inefficiency of African
Water Utilities." Policy Research Working Paper 2890. World Bank, Washington, D.C.
Estache, A., and M. Rossi. 2002. "How Different is the Efficiency of Public and Private Water Compa-
nies." World Bank Economic Review 16(1):13948.
Estache, A,, A. Gomez-Lobo, and D. Leipziger. 2001. "Utilities Privatization and the Poor: Lessons and
Evidence from Latin America." World DevePopment 29(7):1,179-98.
Galiani, S., P. Gertlier, and E. Schargrodsky. 2002. "Water for Life: The Impact of the Privatization of
Water Services on Child Mortality." Working Paper 154. Stanford University, Centre for Research on
Economic Development and Policy Reform, Palo Alto, CA.
Glaeser, E., R. La Porta, F. Lopez-de-Silanes, and A. Shleifer. 2004. "Do Institutions Cause Growth?"
JournaP of Economic Growth 9(3):271-303.
Guasch, J. L. 2004. "Granting and Renegotiating Infrastructure Concessions: Doing It Right." World
Bank, Washington, D.C.
Handley, P. 1997. "A Critical Review of the Build-Operate-Transfer Privatisation Processin Asia." Asian
Journal of Public Administration 19(2):203-43.
Harris, C. 2003. "Private Participation in Infrastructure in Developing Countries: Trends, Impact, and
Policy Lessons." World Bank Working Paper 5. World Bank, Washington, D.C.
Harris, C., J. Hodges, M. Schur, and P. Shukla. 2003. "Infrastructure Projects: A Review of Cancelled
Private Projects." Public Policy for the Private Sector Note 252. World Bank, Private Sector and
Infrastructure Network, Washington, D.C. (http://www.rrr.worldbank.orgPublicPolicyJoumal).
Jalilian, H., C. Kikpatrick, and D. Parker. 2002. "Creating the Conditions for International Business
Expansion: The Impact of Regulation on Economic Growth in Developing Countries." Discussion
Paper 54. University of Manchester, Institute for Development Policy and Management, Centre on
Regulation and Competition, Manchester, UK.
Johnstone, N., and L. Wood. 2001. "Introduction." In N. Johnstone and L. Wood, eds., Private Firms
and Public Water: Realising Social and Environmental Objectives in Developing Countries.
Cheltenham, UK: Edward Elgar.
Kaufmann, D., A. Kraay, and P. Zoido-Lobat6n. 2002. "Governance Matters 11: Updated Indicators for
2000/01." Policy Research Working Paper 2196. World Bank, Washington, D.C.
Kirkpatrick, C., and D. Parker. 2005. "Domestic Regulation and the WTO: The Case of Water Services."
World Economy 28(10):1491-508.
Kodde, D. A., and F. C. Palm. 1986. "Wald Criteria for Jointly Testing Equality and Inequality
Restrictions." Econometrics 54(5):1, 243-8.
McIntosh, A. C. 2003. Asian Water Supplies: Reaching the Urban Poor. London: Asian Development
Bank and International Water Association.
Megginson, W. I.., and J. M. Netter. 2001. "From State to Market: a Survey of Empirical Studies on
Privatization." Journal of Economic Literature 39(2):321-89.
Mitlin, D. 2002. "Competition, Regulation and the Urban Poor: A Case Study of Water." Discussion
Paper 37, University of Manchester, Institute for Development Policy and Management, Centre on
Regulation and Competition, Manchester, UK.
North, D. C. 1990. Institutions, institutional Change, and Economic Performance. Cambridge:
Cambridge University Press.
OECD (Organisationfor EconomicCo-operation and Development). 2003. "Water Partnerships: Striking a
Balance." OECD Observer, April 7. Paris.
Paradi, J. C., and C. Schaffnit. 2004. "Commercial Branch Performance Evaluation and Results Com-
munication in a Canadian Bank-A DEA Application." European Journal of Operational Research
156(3):719-35.
Parker, D., and C. Kirkpatrick. 2005. "Privatisation in Developing Countries: a Review of the Evidence
and the Policy Lessons."Journal of Development Studies 41(4):513-41.
Rivera, D. 1996. "Private Sector Participation in the Water Supply and Wastewater Sector: Lessons from
Six Developing Countries." World Bank, Washington, D.C.
Rodrik, D., A. Subramanian, and F. Trebbi. 2004. "Institutions Rule: The Primacy of Institutions
over Geography and Integration in Economic Development." Journal of Economic Growth
9(2):131-65.
Ruggiero,J. 1996. "On the Measurement of Technical Efficiency in the Public Sector." European Journal
of Operational Research 90(3):553-65.
. 2004. "Performance Evaluation When Non-discretionary Factors Correlate with Technical
Efficiency." European Journal of Operational Research 159(1):250-57.
Shapiro, C., and R. D. Willig. 1990. "Economic Rationales for the Scope of Privatization." In E. N.
Suleiman and J. Waterbury, eds., The Political Economy of Public Sector Reform and Privatization.
Boulder, CO: Westview Press.
Shirley, M., ed. 2002. Thirsting for Efficiency: the Economics and Politics of Urban Water System
Reform. Oxford: Elsevier Press.
Shirley, M., and C. Menard. 2002. "Cities Awash: Reforming Urban Water Systems in Developing
Countries." In M. Shirley, ed., Thirsting for Efficiency: the Economics and Politics of Urban Water
System Reform. Oxford: Elsevier Press.
Snell,S. 1998. "Water and Sanitation Services for the Urban Poor-Small-Scale Providers: Typology and
Profiles." United Nations Development Programme-World Bank Water and Sanitation Program,
Washington, D.C.
Spiller, P. T., and W. Savedoff.1999. "Government Opportunism and the Provision of Water." In W. D.
Savedoffand P. T. Spiller, eds., Spilled Water: an Institutional Commitment to the Provision of Water
Services. Washington, D.C.: Inter-American Development Bank.
Stevenson, R. 1980. "Likelihood Functions for Generalized Stochastic Frontier Estimation." Journal of
Econometrics 13(1):57-66.
Tamayo, G., R. Barrantes, E. Conterno, and A. Bustamente. 1999. "Reform Efforts and Low-Level
Equilibrium in the Peruvian Water Sector." In W. D. Savedoffand P. T. Spiller, eds., Spilled Water: an
Institutional Commitment to the Provision of Water Services. Washington, D.C.: Inter-American
Development Bank.
TEN(TheEnergy and ResourcesInstitute). 2003. "Centre on Regulation and Competition Draft Project."
Report 2003PP44. The Energy and Resources Institute, New Delhi.
Wang, H., and P. Schmidt. 2002. "One-Step and Two-Step Estimation of the Effects of Exogenous
Variables on Technical Efficiency Levels." Journal of Productivity Analysis 18(2):120-44.
Williamson, 0. E. 1985. The Economic Institutions of Capitalism: Firms, Markets, and History.
Cambridge: Cambridge University Press.
World Bank. 1995. Bureaucrats in Business: the Economics and Politics of Government Ownership.
Oxford: Oxford University Press.
-. 2002. Word Development Indicators Online. Washington, D.C.
Kirkpatrick, Parker, i;md Zhang 163
-. 2003. Private Participation in Infrastructure: Trends in Developing Countries in 1990-2001.
Washington, D.C.
. 2004a. Doing Business in 2004: Understanding Regulation. World Bank, Oxford University
Press: Washington, D.C.
. 2004b. Reforming Infrastmcture: Prit,atization, Regulation, and Competition. World Bank
Policy Research Report. Oxford: Oxford University Press.
WRI (World Resources Institute). 2003. "Earth Trends." World Resources Data CD. Washington, D.C.
Zhang, Y.-F., C. Kirkpatrick, and D. Parker. 2003. "ElectricitySector Reform in Developing Countries:
An Econometric Assessment of the Effects of Privatisation, Competition and Regulation." Discussion
Paper 31. University of Manchester, Institute for Development Policy and Management, Centre on
Regulation and Competition, Manchester, UK.
ECONOMIES
ED$-TED MARCEL FAFGHAMPS,
BY ERNE% ARYEETE~
MICHAEL SLEANEY AND AUGUSTINFOSU
The Journal of African Economiesis a vehicle to carry
rigorous economic analysis,focused entirely on Africa,
for Africans and anyone interested in the continent -
be they consultants, policymakers, academics, traders,
i financiers, developmcnt agents or aid workers.
1 Each issue of thejournal of Afncan Oonomiescontains
I applied research of the highcst standard, together
with a comprehensive book review section, and a listing
of the current working papers from around the world.
N O W W I T H ADVANCE ACCESS
Advance Access enables us to publish papers online
soon after they have been accepted for publication
and well ahead of their appearance in the printed
journal, thus greatly reducing publication times and
ensuring our readers stay in touch with the very latest
deveIopments in the field. See our website for more
details.
J A E ONLINE
Visit theJo.umal of AfFicarn EconomiPswebsite to read a
free online sample issue and find articles of interest
by searching across the contents of past issues using
keywords and author's names. The website also
contains advance access papers, subscription prices,
submission information, and our free content alerting
service.
OXFORD
UNIVERSITY PRESS
-
Forthcomingpapers in I
I
I
I
I
I
I 1
THE WORLD BANK
ECONOMIC REVIEW
Volume 20, Number 2,2006
PREFERENCEEROSION: ASYMPOSIUM
Introduction
Bernard Hoekman
Doha MerchandiseTrade Reform:What's at Stake for Developing
Countries?
Kym Anderson, Will Martin, and Dominique van der Mensbruggbe
Ill I 111
Preference Erosion and MultilateralTrade Liberalization
Joseph E Francois, Bernard Hoekman, and Miriam Manchin
Trade Preferencesto Small DevelopingCountries and the Welfare
Costs of Lost Multilateral Liberalization
Nuno Lim6o and Marcelo Olarreaga
I
PriceEffects of PreferentialMarket Access: Caribbean Basin ~nitiatkve
and the ApparelSector
Caglar 0zden and Gunjan Sharma
Aid and the SupplySide: PublicInvestment, Export Performance
and Dutch Disease in Low Income Countries
Christopher S. Adam and David L. Bevan
III Infrastructure, Externalities, And EconomicDevelopment:A studL III