THE WORLD BANK ECONOMIC REVIEW
Volume 17 2003 Number 1
Grandmothers and Granddaughters: Old Age Pension and
Intrahousehold Allocation in South Africa
Esther D d o
Public Policy and Extended Families: Evidence from Pensions
in South Africa
Marianne Bertrand, Sendhil hlullainathan, and DouglasIlliller
Economic, Demographic, and Institutional Determinants of Life
Insurance Consumption across Countries
norsten Beck andIan Webb
Benefits on the Margin: Observations on Marginal Benefit Incidence 89
StephenD. Younger
Reducing Child Malnutrition: How Far Does Income
Growth Take us?
LawrenceHaddad, HaroldAlderman, SimonAppleton, Lina Song,
and Eisehac Yohannes
A NEWDEVELOPMENT DATABASE
Particularism around the World
Jessica Seddon Walack,Alejandro Gaviria, UgoPanizza,
and Emesto Stein
The WorldBBanEcon~~cReviewavailable online at no extra cost with aprint
is
subscription in 2003.Activateyour subscriptionat http://www3.oup.co.uk/Register.
Keep up-to-date with the latest contents of TheWorldBankEconomicReview
by registeringfor our eTOC service at http://www3.oup.co.uk/jnls/tocmail.
This serviceis freely availableto all, no subscriptionrequired.
THE WORLD BANK
ECONOMIC REVIEW
EDITOR
Franqois Bourguigrlon, World Bank
EDITORIAL BOARD
Abhijit Banerjee, MassachusettsInstitute of Ravi Kanbur, Cornell University, USA
techno lo^, USA Elizabeth M. King, WorldBank
Kaushik Basu, Cornell University, USA Justin Yifu Lin, China Centerfor Economic
T i m Besley, London SchoolfEconomics, UK Research, Peking Univmsity, China
Anne Case, Prinbton University, USA Mustapha Kame1 Nabli, WorldBank
Stijn Claessens, UniversityofAmsterdam, Juan Pablo Nicolini, Universidaddi Tella,
TheNetherlands Argentina
Paul Collier, WorldBank Howard Pack, UniversityofPennsylvania, USA
David R. Dollar, WorldBank Jean-Philippe Platteau, Faculte's Universitaires
Antonio Estache, WorldBank Notre-Dame de la Paix, Belgium
Augustin Kwasi Fosu, Afiican Economic Boris Pleskovic, WorldBank
Research Council,Kenya Martin RavaUion, WorldBank
Mark Gersovitz, TheJohns Hopkins Carmen Reinhart, UniversityofMayland, USA
University, USA Mark R. Rosennveig, Harvard University,USA .
Jan W i e m Gunning, Free University, Joseph E. Stiglitz, Columbia University, USA
Amsterdam, TheNetherlands Moshe Syrquin, University $Miami, USA
Jeffrey S. Hammer, WorldBank Vinod Thomas, WorldBank
Karla Hoff, WorldBank I,. Alan Winters, University ofsussex, UK
The WorIa'BankEconomicReview is aprofessionaljournal for the dissemination ofWorld Bank-sponsored
and outside research that mayinform policy analysesand choices. It is directed to an international readership
among economists and social scientists in government, business, and international agencies, as well as in
universities and development research institutions. The Review emphasizespolicy 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 will illustrate how professional research can shed light on policy choices.
-
Inconsistency with Bank policywill not be grounds for rejection of an article.
Articles will be drawn from work conducted by World Bank staff and consultants and from papers submitted
by outside researchers. Before being accepted for publication, all articleswill be reviewedby two refereeswho
are not members of the Bank's staffand one World Bank staff member. Articles must also be recommended
by a member of the Editorial Board. Non-Bank contributors are requested to submit a proposal of not more
than two pages in length to the Editor or a member of the Editorial Board before sending in their paper.
Comments or brief notes responding to Review articlesarewelcome and willbe considered forpublication to
the extent that space ~ermits.Pleasedirect alleditorialcorrespondencetothe Editor, The WorldBunk Elonomic
Review,The World Bank, 1818 H Street, Washington, DC 20433, USA, or wber@Worldbank.org.
For more information, $ease visit the Web sites of the Economic Review at
www.wber.oupjournals.org, the World Bank at www.worldbank.org, and
Oxford University Press at www.oup-usa.org.
The WorldBankEconomicReviav(1SSN 0258-6770) is published three times a year by Oxford University
Press, 2001 Evans Road, Cary, NC 27513-2009 for The International Bank for Reconstruction and
Development / THE WORLD BANK. Communications regarding original articles and editorial management
should be addressed toThe Editor, TheWorldBankEconomicReviav,66, avenue d'Iena, 75116 Paris, France.
E-mail: wber&orldbank.org.
Oxford University Press is a department of the University of Oxford. It furthers the University's objective
of excellence in research, scholarship, and education by publishing worldwide.
SUBSCRIPTIONS: Subscriptionis on a yearly basis. The annual rates are US140 (E30in UK and Europe) for
individuals;USIlOl (L71 in UK and Europe) for academic libraries; US96123 (L84 in UK and Europe) for
corporations. Single issues are availablefor UStl6 (El2 in UK and Europe) for individuals;US143 (�30 in
UK and Europe) for academic libraries;US9652 (L35 in UK and Europe) for corporations. Allprices include
postage. Individual rates are applicable onlywhen a subscriptionis for individual use and are not availableif
delivery is made to a corporate address. Subsmpuonsareprotided hee ofcharge to non-OECD countries. All
subscription requests, singleissue orders, changes of address, and claimsfor missing issues should be sent to:
NorthAmerica: Oxford U~liversityPress,Journals Customer Service,2001 Evans Road, Cary, NC 27513-
2009, USA. Toll-free in the USA and Canada: 800-852-7323, or 919-677-0977. Fax: 919-677-1714.
E-mail: jnlorders@oup-usa.org.
Elsewhere: Oxford University Press,Journals SubscriptionsDepartment, Great Clarendon Street, Oxford
OX2 6DP, UK. Tel: +44 1865 353907. Fax: +44 1865 353485. E-mail: jnls.cust.serv@oup.co.uk.
ADVERTISING: Helen Pearson, Oxford Journals Advertising, P.O. Box 347, Abingdon SO, OX14 lGJ,
UK. TeVFax: +44 1235 201904. E-mail: helen@oxfordads.com.
BACK ISSUES:The current plus two back volumes are available from Oxford University Press. Prcvious
volumes (from 1997) can be obtained from Periodicals Service Company, 11Main Street, Germantown,
NY 12526, USA. Tel: 518-537-4700.
REQUESTS FOR PERMISSIONS,REPRINTS, AKD PHOTOCOPIES: All rights reserved; no part of this publica-
tion may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, elec-
tronic, mechanical, photocopying, recording, or otherwise,without either prior written permission of the
publisher (Oxford UniversityPress,Journals Rights and Permissions, Great Clarendon Street, Oxford OX2
6DP, UK; tel: +44 1865 354490; fax: +44 1865 353485) or a license permitting restricted copying issued
in the USA by the Copyright Clearance Center, 222 Rosewood Drive, Danvers, MA 01923 (fax: 978-
646-8600), or in the UK by the Copyright Licensing Agency Ltd., 90 Tottenham Court Road, London
W I T 4LP, UK (fax: 020 7631 5500). Reprints of individual articles are available only from the authors.
COPYRIGHT: Copyright @ 2003The International Bank for Reconst~ctionand Development / THE WORLD
BANK. It is a condition of publication in the journal that authors grant an exclusive license to The Interna-
tional Rankfor Reconstmction and Development / THE WORLD BANK.However, requests for permission to
reprint material found in the journal should come to Oxford University Press. This ensures that requests
from third parties to reproduce articles are handled efficiently and consistently and will also allow the ar-
ticle to be as widely disseminated as possible.As part of the license agreement, authors may use their own
material in other publications provided that the journal is acknowledged as the original place of publica-
tion and Oxford University Press as the publisher.
INDEXING AND ABSTRACTING:The WorldBank Economic Review is indexed and/or abstracted by GIB
Abshacts, Current ContentdSocial and Behavioral Sciences,Journal ofEconomic LiteraturdEconLit, PAIS
International, RePEc (Researchin EconomicPapem), and SocialSciencesCihtionIndex. The microform edi-
tion is available through ProQuest (formerly UMI), 300 North Zceb Road, Ann Arbor, M I 48106, USA.
PAPEK 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).
POSTAL INFORMATION: The WorldBankEconomicRevim (ISSN 0258-6770) is published three times a
year by Oxford University Press, 2001 Evans Road, Cary, NC 27513-2009. Send address changes to The
WoddBankEconomicRevjew,Journals Customer ServiceDepartment, Oxford UniversityPress,2001 Evans
Road, Cary, NC 27513-2009.
the world bank economic review, vol. 17, no. 1 1�25
Grandmothers and Granddaughters:
Old-Age Pensions and Intrahousehold
Allocation in South Africa
Esther Duflo
This article evaluates the impact of a large cash transfer program in South Africa on
children's nutritional status and investigates whether the gender of the recipient affects
that impact. In the early 1990s the benefits and coverage of the South African social
pension program were expanded for the black population. In 1993 the benefits were
about twice the median per capita income in rural areas. More than a quarter of black
South African children under age five live with a pension recipient. Estimates suggest
that pensions received by women had a large impact on the anthropometric status
(weight for height and height for age) of girls but little effect on that of boys. No simi-
lar effect is found for pensions received by men. This suggests that the efficiency of
public transfer programs may depend on the gender of the recipient.
Cash transfers are still rare in developing economies. But they are being proposed
more often by policymakers and academics as a viable way to redistribute resources.
Proponents argue that improvements in the ability to handle cash transfers have
made such transactions much easier to implement on a large scale--and less prone
to corruption--than in-kind benefits (such as free health or education services).
Others worry that redistributing money to adults may be less efficient than sub-
sidizing investment in children if parents do not fully internalize the returns to
investing in child health.
There is evidence that inadequate nutrition in childhood affects long-term
physical development as well as the development of cognitive skills (Barker
1990).1 This in turn affects productivity later in life (Dasgupta 1993; Strauss and
Thomas 1998; Schultz 1999). Low levels of investment in child health therefore
have far-reaching consequences for economic growth, distribution, and welfare.
Esther Duflo is Professor of Economics at the Massachusetts Institute of Technology. Her e-mail
address is eduflo@mit.edu. The author gratefully acknowledges financial suppport from Fondation Thiers
and the Alfred P. Sloan Foundation. The author thanks Josh Angrist, Abhijit Banerjee, Tim Besley,
Fran�ois Bourguignon, Anne Case, Pierre-Andr� Chiappori, Angus Deaton, Andrew Foster, Robert
Jensen, Michael Kremer, Emmanuel Saez, Duncan Thomas, and three referees for useful comments.
1. Balazs and others (1986) review the biomedical and empirical literature on the relationship be-
tween early childhood nutrition and the development of intelligence. Miguel and Kremer (2001) show
that school attendance is higher among children treated for worms.
DOI: 10.1093/wber/lhg013
� 2003 The International Bank for Reconstruction and Development / THE WORLD BANK
1
2 the world bank economic review, vol. 17, no. 1
But cash transfers may result in improvements in the consumption of adults but
not in children's human capital, even if investments in child health are ineffi-
ciently low.
This debate is closely linked to questions about the optimal design of cash
transfers. First, should they be made conditional or unconditional? Evidence from
the United States suggests that in-kind transfers (which are a particular kind of
conditional transfer)--such as the nutrition supplements distributed under the
Special Supplemental Nutrition Program for Women, Infants, and Children (com-
monly known as the wic Program)--are associated with greater improvements
in children's health than are cash transfers--such as Aid to Families with De-
pendent Children (Mayer 1997; Currie 1995). Second, do the characteristics of
the beneficiaries within the household matter? A growing literature suggests that
income or assets in the hands of women are associated with larger improvements
in child health and larger shares of household spending on nutrients, health, and
housing than are resources in the hands of men (Thomas 1990, 1994). Based on
this and similar evidence, policymakers have favored transfers targeted to women.
A prime example of a program combining these two features is Progresa, a pro-
gram implemented in Mexico and replicated in several other Latin American
countries. Payments are made to women conditional on their children attending
school and on their participation in a health care monitoring and food supple-
mentation program. The program has been shown to have significant effects on
children's health, nutrition, and education (Gertler and Boyce 2001; Schultz
2000).
Still, there is little evidence that the gender of the recipient affects the impact
of a cash transfer program.2 The evidence that income in the hands of women is
associated with different expenditures than income in the hands of men is sug-
gestive but could be misleading. Families in which women work or own more
assets could differ in many respects from families in which women have no ac-
cess to resources and thus make different decisions. Though the evidence on
Progresa suggests that conditional transfers to women can work, it does not
answer two other questions: whether unconditional cash transfers can have
positive effects and whether these effects are sensitive to the gender of the recipi-
ent (because all the recipients were women).
In this article I seek to answer these two questions. I evaluate the impact of
the South African old-age pension program (one of the few successful cash transfer
programs in the developing world) and compare its effects by gender of the pen-
sion recipient using data from a 1993 national household survey. Historically
the program was racially discriminatory. At the end of the apartheid era, the
2. The study by Lundberg and others (1996) is an exception. The authors investigate the effects of
a change in the mode of allocation of child benefits in the United Kingdom from a tax credit to a direct
payment to the mother. This transfer "from the wallet to the purse" appears to have been associated
with an increase in the consumption of women's and children's clothing relative to men's clothing in
households with children.
Duflo 3
government made a commitment to achieving parity in benefits and eligibility
requirements for whites and Africans.3 Parity was achieved mostly by increas-
ing the benefits received by Africans. The new system is universal and noncon-
tributory. All women over the age of 60 and all men over age 65 are entitled to
benefits, subject to a means test. More than one member of a household can
receive the pension at the same time.
In 1993, 80 percent of African women over age 60 and 77 percent of African
men over age 65 received the pension. Most received the maximum of 370 rand
(R) a month, roughly twice the median per capita income in rural areas. More
than a quarter of African children under the age of five live with a pension recipi-
ent, because grandparents often live in extended households with their children
and grandchildren.4 The old-age pension program thus provides an opportunity
to evaluate the effect on child nutritional status of an unusually large income
transfer that was not targeted specifically to either men or women but could be
received by both.
I investigate the effect of men's and women's pensions on child nutrition as
reflected in anthropometric indicators--weight for height and height for age. The
identification of this effect is complicated by the fact that children living with a
pension recipient are relatively disadvantaged on average. Case and Deaton
(1998) have shown that the program was effective in transferring money pre-
dominantly to poor households, especially to households with poor children.
South Africa began to expand the pension program at the end of 1991, and in
1993 the program had been fully operating in all areas for less than a year. Not
surprisingly then, because child height in 1993 reflected past as well as current
nutrition, children living with a pension recipient were on average smaller for
their age than were other children.
To address this problem, I first make use of the fact that pension receipt ex-
hibits a discontinuity at age 60 for women and age 65 for men. Unlike height for
age, children's weight for height responds quickly to changes in the environment.
I compare the weight for height of children in households with no member eli-
gible for the pension, those in households with an eligible man, and those in
households with an eligible woman after controlling for the presence of a man
or a woman who is old but not old enough to be eligible (for example, a woman
between 55 and 60). The difference is then normalized by the difference in the
probability to receive the pension across these two groups. The results suggest
that the pensions received by women increased the weight for height of girls by
1.19 standard deviations but did not significantly increase that of boys. Pensions
received by men are not associated with an improvement in the nutritional sta-
tus of either girls or boys.
3. I generally use the official terms for racial groups in South Africa (Africans, whites, colored, and
Indians).
4. These living arrangements are due in large part to apartheid rules, which prohibited the families
of migrant workers--those working in the mines or as domestic servants--from joining them.
4 the world bank economic review, vol. 17, no. 1
This comparison would be misleading if there were intrinsic differences be-
tween households with a member between age 55 and 60 and households with
a member over age 60 or if the expansion of the pension program had led to
endogenous changes in the composition of households. Thus, in a second step I
make use of the fact that height for age reflects past as well as current nutrition
and illnesses. Because all children were measured around the same date, if the
pension indeed affects nutrition, older children would have had longer periods
of inadequate nutrition. Thus, older children in eligible households should be
smaller than those in noneligible households. But for younger children the dif-
ference between those in eligible households and those in noneligible households
should be reduced or even reversed. The basic idea of the identification strategy
is thus to estimate whether the relative disadvantage in height between children
in eligible and those in noneligible households are smaller for younger children
than for older children.5 The results obtained using this second strategy are strik-
ingly similar to those for weight for height: pensions received by women are
associated with an increase of 1.16 standard deviations in the height for age of
girls but had no significant effect on that of boys. Pensions received by men are
not associated with an improvement in the height for age of either boys or girls.
I. The South African Old-Age Pension Program
This section presents a brief history and overview of the South African old-age
pension program, drawing extensively on Van der Berg (1994), Lund (1993),
and Case and Deaton (1998), as well as descriptive statistics on the program.
Description of the Program
South Africa first introduced social pensions in the 1920s for whites, mainly as
a social safety net for the minority of white workers not covered by occupational
pensions. The pensions were gradually extended, but with very dissimilar bene-
fit levels, to other racial groups. During the apartheid era the system was racially
discriminatory in several respects. First, different means tests were applied to each
racial group. In 1984, for example, benefits were withdrawn for incomes larger
than R700 a year for Africans but for incomes larger than R2250 a year for
whites. Second, the benefit levels were different. In the early 1980s benefits for
whites were 10 times those for Africans.6 Third, the delivery systems were dif-
ferent. Pensions for whites were distributed through postal offices, whereas those
for Africans were distributed through mobile pay points that did not reach very
far into rural areas. Finally, officials often intentionally underestimated people's
ages, removed people from the computer lists, or otherwise limited the access of
legally eligible Africans to reduce the cost of pensions (Lund 1993).
5. In an earlier work I proposed a nonparametric version of this test (Duflo 2000a).
6. The nonpension incomes of Africans were also much smaller, so as a share of income the differ-
ence was much smaller (Van der Berg 1994).
Duflo 5
Pressures for equity in the treatment of racial groups became strong toward
the end of apartheid, and in 1989 the government made a commitment to achiev-
ing racial parity in the pension program (Van der Berg 1994). Extending the social
pension to the entire population took several years, and the program was fully
operating in all areas only at the beginning of 1993. The benefits for Africans
rose gradually in the 1980s--from R1555 a year in 1980 to R2096 in 1990 (both
in 1990 rands)--whereas those for whites declined rapidly. The benefits for
Africans increased much faster in the 1990s--to R2444 in 1991, to R2677 in
1992, and to R3081 in 1993 (all in 1990 rands). Monthly benefits in 1993 were
R370 (1993 rands), and the monthly per capita household income of Africans
in the sample averaged R149. Because of the high unemployment in South Africa,
pension recipients are often the main income earner in their household.
In 1992 the means test was modified and unified across races. The current
system is universal and noncontributory. Payments are made to women over age
60 and to men over age 65, subject to a means test. In calculating the means test,
a couple's resources are roughly divided by two and the income of other house-
hold members is not taken into account. The pension program therefore pro-
vides no direct incentives to partition the household or for other household
members to stop working. In practice, the means test does not seem to be ap-
plied very finely. It is mainly effective in excluding most whites as well as Afri-
cans with a private pension.
In 1993, 60 percent of men and 77 percent of women in the sample who were
eligible on the basis of their age were receiving a pension (table 1). Of these, most
received the maximum amount. There is no good estimate of the coverage in
earlier periods for two reasons. First, social pensions were administered by sev-
eral different agencies, which made any evaluation difficult. Second, surveys (in-
cluding the 1991 census) excluded the independent homelands, where many
Africans live. The coverage increased substantially in the 1990s as a result of a
new attitude in the program administration, a modification of the means test,
computerization of the system, and substantial improvements in the delivery
system.7
Data and Descriptive Statistics
The data come from a national survey carried out jointly by the World Bank
and the South African Labor and Development Research Unit at the University
of Cape Town. During the last five months of 1993, 9000 randomly selected
households of all races and in all areas were interviewed. As part of the survey
the height and weight of all children under seven years old were measured. Be-
cause environmental factors are especially important determinants of height in
early childhood, the World Health Organization (who) recommends limiting
7. For example, in the province of KwaZulu Natal the pension is distributed once or twice a month
through mobile pay points equipped with automated teller machines that have a fingerprint recogni-
tion system (Case and Deaton 1998).
6 the world bank economic review, vol. 17, no. 1
Table 1. Probability of Receiving the Old-Age Pension by Age
and Gender and Share of Children in the Survey Living in
Households with Adults in this Age and Gender Group, 1993
(percent)
Share of age group Share of children living
receiving pension with age group members
Men by age (years)
50�54 2.8 9.8
55�59 4.7 7.6
60�64 22.0 5.5
65 and over 60.0 8.0
Women by age (years)
50�54 13.6 8.2
55�59 16.4 10.9
60 and over 77.0 21.0
Source: Author's calculation from the 1993 saldru national household survey.
the analysis of height and weight measures to children ages zero to five years
(who 1986). Moreover, there appears to have been problems in the survey with
the measurement of the oldest children.8 For these reasons I follow earlier stud-
ies and restrict the sample to children ages 6�60 months (Case and Deaton 1998;
Le Roux 1995). For each age in months I construct height-for-age z-scores by
subtracting the median and dividing by the standard error in the corresponding
age and sex group in the reference population established by the U.S. National
Center for Health Statistics (a group of well-nourished U.S. children). I construct
weight-for-height z-scores in a similar way.9
Households with either a woman or a man eligible for the pension have simi-
lar characteristics (table 2). But compared with households with no eligible
member, these households are poorer even after pension income is included. Not
surprisingly, they are often characterized by the presence of a grandparent and
the absence of the child's father (67 percent) or mother (18 percent). They are
also more likely to live in a rural area.
Children in households with an eligible member are smaller than other chil-
dren. This is not surprising. Even if the greater coverage and benefits of the pen-
sion program had led to better child nutrition, height for age still reflects past
deprivations or illnesses, especially among older children. In contrast, average
weight for height, a measure of short-run nutritional status, is higher in house-
holds with an eligible woman than in those with an eligible man or with no one
who is eligible.
8. Some six- and seven-year-old children were recorded as eight by the interviewers and thus not
measured. It seems likely that if a child was tall the interviewer would have assumed that the child was
older and therefore mistakenly excluded that child.
9. This normalization does not affect the analysis, which relies on the comparison of the height of
children in eligible households with that of children of the same age in noneligible households and controls
for the child's age.
Duflo 7
Table 2. Descriptive Statistics
Eligibility for pension
Woman Man None
Household characteristics
Mother's education 5.70 5.78 5.17
(0.16) (0.24) (0.086)
Father's education 5.07 4.20 4.54
(0.27) (0.46) ( 0.11)
Rural residence 0.75 0.83 0.67
(0.018) (0.028) (0.012)
Grandparent in household 0.95 0.89 0.42
(0.0081) (0.021) (0.012)
Father is absent 0.67 0.66 0.41
(0.020) (0.033) (0.012)
Mother is absent 0.18 0.14 0.08
(0.016) (0.023) (0.0059)
Household size 10.5 10.5 7.6
(0.21) (0.30) (0.086)
Income and pension receipt
Man receives pension 0.17 0.68 0.03
(0.016) (0.034) (0.0041)
Woman receives pension 0.79 0.47 0.04
(0.018) (0.037) (0.0050)
Nonpension income 723 637 908
(36) (51) (22)
Pension income 325 389 23
(9.6) (20) (2.2)
Per capita income 121 123 149
(4.5) (7.3) (3.9)
Anthropometric data
Height-for-age z-score �1.38 �1.46 �1.21
(0.072) (0.13) (0.036)
Weight-for-height z-score 0.28 0.12 0.15
(0.08) (0.15) (0.04)
Observations 816 286 2,380
Note: Standard errors are in parentheses.
Source: Author's calculations.
This suggests that pensions received by women may have led to an improve-
ment in children's health, whereas pensions received by men had no comparable
effect. The next two sections elaborate on this evidence.
II. Effect of the Pension on Weight for Height
This section presents estimates of the effect of the pension on weight for height,
a measure of short-run nutrition and illness.
8 the world bank economic review, vol. 17, no. 1
Empirical Specification
The weight for height of children reflects short-run nutrition and illnesses and
recovers quickly after periods of malnutrition when proper nutrition is resumed
(Ashworth 1969; Martorell and Habicht 1986). Thus, it reflects the impact of
current nutrition decisions by parents as well as that of the environment.
Comparing the 1993 weight for height of children living with an eligible
woman, an eligible man, or no eligible household member would confound the
effect of pension eligibility with the effect of differences in background. To con-
trol for these differences, I estimate the effect of having an eligible man or an
eligible woman in the household after controlling for the presence of a man or a
woman over age 50, a man or a woman over age 55, and a man over age 60 (in
1992) as well as a series of household-specific control variables described later.10
Some people who are not yet age-eligible receive the pension, but the probabil-
ity of receiving the pension does increase discretely at age 60 for women and age
65 for men (see table 1). At the time of the survey the pension program was widely
known, so those who were close in age to being eligible for the pension expected
to receive it. With the presence of someone close to being eligible controlled for,
a positive coefficient of the eligibility dummy variable must therefore indicate
the presence of credit constraints. Moreover, even if there are credit constraints,
the weight given to the preferences of a woman over age 55 may reflect the fact
that she will earn a pension when she turns 60. To the extent that this is true,
the difference between the coefficient of a man's eligibility and that of a woman's
eligibility in this specification is an underestimate of the difference between the
effect of money given to men and that of money given to women.
The regression estimated is therefore:
4
(1) wijk = pfEf + pmEm + gl1( l = k)+ Wijkl + Xijkd + wijk
j=1
where wijk is the weight-for-height z-score of a child born in cohort k in family
f, Ef is equal to one if there is an eligible woman in the household and zero other-
wise, Em is equal to one if there is an eligible man in the household and zero
otherwise, and 1( l = k)is a dummy variable indicating the year of birth of the child.
Wijk is a vector of variables indicating whether there is a woman over age 51 in
the household, a man over age 50, a woman over age 56, a man over age 56, and
a man over age 61.11 Xijk is a vector of family background variables: mother's
and father's education levels; rural, urban, or metropolitan residence; mother's
and father's ages; size of household; and the number of household members in
the age categories 0�5, 6�14, 15�24, and 25�49 years.12
10. This strategy was used by Case and Deaton (1998), Bertrand and others (1999), and Edmonds
and others (2001).
11. Thus, these individuals were respectively age 49, 55, and 60 in 1992.
12. I have replaced the relevant variables with sample means where the father or mother of the child
was absent.
Duflo 9
Table 3. Effect of the Old-Age Pension Program on Weight for Height: ols
and 2sls Regressions
ols 2sls
Variable (1) (2) (3) (4) (5) (6) (7)
Girls
Eligible household 0.14 0.35* 0.34*
(0.12) (0.17) (0.17)
Woman eligiblea 0.24* 0.61* 0.61* 1.19*
(0.12) (0.19) (0.19) (0.41)
Man eligibleb �0.011 0.11 0.056 �0.097
(0.22) (0.28) (0.19) (0.74)
Observations 1574 1574 1533 1574 1574 1533 1533
Boys
Eligible household 0.0012 0.022 0.030
(0.13) (0.22) (0.24)
Woman eligiblea 0.066 0.28 0.31 0.58
(0.14) (0.28) (0.28) (0.53)
Man eligibleb �0.059 �0.25 �0.25 �0.69
(0.22) (0.34) (0.35) (0.91)
Observations 1670 1670 1627 1670 1670 1627 1627
Control variables
Presence of older membersc No Yes Yes No Yes Yes Yes
Family background variablesd No No Yes No No Yes Yes
Child age dummy variablese Yes Yes Yes Yes Yes Yes Yes
*Significant at the 5 percent level.
Note: The instruments in column 7 are woman eligible and man eligible (the first stage is in table A-1).
Standard errors (robust to correlation of residuals within households and heteroscedasticity) are in
parentheses.
aIn column 7 this variable is replaced by a dummy for whether a woman receives the pension.
bIn column 7 this variable is replaced by a dummy for whether a man receives the pension.
cPresence of a woman over age 50, a man over age 50, a woman over age 56, a man over age 56,
and a man over age 61.
dFather's age and education; mother's age and education; rural or metropolitan residence (urban is
the omitted category); size of household; and number of members ages 0�5, 6�14, 15�24, and 25�49.
eDummy variables for whether the child was born in 1991, 1990, or 1989.
Source: Author's calculations.
The identification assumption underlying this exercise is that there is no sys-
tematic difference in nutrition between eligible and noneligible households with
an elderly member. As I discuss later, this assumption may be problematic, and
I present results for an alternative specification that relaxes it.
Results
The results from estimating equation 1 are presented in table 3. Columns 1�3
do not distinguish by gender of the eligible household member. For girls the
coefficient is positive but insignificant without controlling for the presence of
noneligible members over age 50 (column 1). When these controls are introduced,
the coefficient more than doubles (0.35) and becomes significant (column 2).
10 the world bank economic review, vol. 17, no. 1
Introducing additional control variables does not affect the coefficient (column
3). As expected, because extended families tend to be poorer than nuclear fami-
lies, the direct effect of having household members over age 56 (not reported
here) is negative and similar across genders. The similarity between the coeffi-
cients in columns 2 and 3 is reassuring. It suggests that the dummy variables for
the presence of elderly members capture the effect of observed (and hopefully
unobserved) family variables. For boys there seems to be no effect from having
an eligible household member, regardless of whether the presence of other elderly
members in the household is controlled for.
Columns 4�6 estimate the effects separately according to the gender of the
eligible household member. For girls column 6 suggests that having a woman
eligible increases weight for height by 0.6 standard deviation (with a standard
error of 0.19). In contrast, having a man eligible increases weight for height by
an insignificant 0.056 standard deviation. The coefficients on men's and women's
eligibility differ statistically from each other. (The F-statistic is 2.50, with a p-value
of 0.11.) For boys we can now detect a smaller (0.28 standard deviation) and
insignificant positive effect of a woman's eligibility on weight for height and a
negative (and insignificant) effect of a man's eligibility. Again, the coefficients
are very similar in columns 5 and 6.
Column 7 checks that the difference between the coefficient of a man's pen-
sion income and that of a woman's pension income is not simply due to the fact
that, conditional on being age-eligible, men claim the pension much less often
than women (perhaps because they are more likely to have worked and to have
a private pension). To do this, I estimate the following relationship using two-
stage least squares (2sls):
4
(2) wihk = afPENSf + amPENSm + gj1( j = k)+ Wihkl + Xihkd + wihk
j=1
where PENSf is a dummy variable indicating whether a woman in the household
receives a pension and PENSm a dummy variable indicating whether a man in the
household receives a pension. The instruments are dummy variables for the pres-
ence of an eligible woman and the presence of an eligible man. Not surprisingly,
the first stage is strong (table A-1, columns 1 and 2). The coefficient of the eligible
man dummy variable in the regression predicting a male pension is 0.39 (with a
t-statistic of 8), and that of the eligible woman dummy variable in the regression
predicting a female pension is 0.55 (with a t-statistic of more than 9).13
The results in column 7 of table 3 confirm that the differences between the
effects of men's and women's pension eligibility found in columns 5 and 6 are
not an artifact of the smaller first stage for a man's pension. For girls these esti-
mates suggest that a woman's pension increases weight for height by 1.19 stan-
dard deviations, whereas a man's pension has a small, negative, insignificant
13. These coefficients reflect the difference between the probability of receiving the pension when
eligible and that of receiving the pension when close to eligibility.
Duflo 11
Table 4. Effect of Pension Eligibility on Weight for
Height by Gender of the Intermediate Generation:
ols Regressions
Variable Girls Boys
Mother's mother eligible 0.48* 0.099
(0.21) (0.27)
Father's mother eligible 0.15 0.29
(0.25) (0.30)
Mother's father eligible 0.097 0.00052
(0.34) (0.43)
Father's father eligible 0.22 0.25
(0.48) (0.44)
Observations 1457 1552
Control variables
Presence of older membersa Yes Yes
Family background variablesb Yes Yes
Age dummy variablesc Yes Yes
*Significant at the 5 percent level.
Note: Standard errors (robust to correlation of residuals within house-
holds and heteroscedasticity) are in parentheses.
aDummy variables for whether there is a woman over age 50, a man over
age 50, a woman over age 56, a man over age 56, and a man over age 61.
bFather's age and education; mother's age and education; rural or metro-
politan residence; size of household; and number of members ages 0�5, 6�
14, 15�24, and 25�49.
cDummy variables for whether the child was born in 1991, 1990, or 1989.
Source: Author's calculations.
effect. For boys the coefficient of a woman's pension is positive but only half as
large as the effect for girls (0.58) and insignificant (the standard error is 0.53).
The coefficient of a man's pension is negative (�0.69) and insignificant.
I also examine whether the gender of the parent whose own parent is eligible
has an effect as well. Strikingly, only the eligibility of the mother's mother has a
significant effect on girl's weight for height (table 4).
These results provide some suggestive evidence that the old-age pension had
very different effects on child health depending on whether it was received by a
woman or by a man. Moreover, there appears to be an all-female link, because
the pension seems to be effective only if received by the mother of a girl's mother.
There are caveats to this interpretation, however, which I now discuss.
Can Unobserved Differences between Eligible and Noneligible
Households Explain the Results?
The essential difficulty is whether controlling for the presence of members over
age 50, 56, and 61 adequately controls for the differences between eligible and
noneligible households. A first problem could be differences between households
with a member over age 55 and those with a member over age 60. For example,
conditional on a household's having three generations, the presence of an elderly
grandparent may be a sign of a relatively healthy household. That a grandmother
12 the world bank economic review, vol. 17, no. 1
is old indicates that she, the mother, or both did not have children very early. (It
could also indicate that the grandmother had many children, but I control di-
rectly for this.) That the grandmother is old but still alive could indicate that
household members are generally healthy. Both effects would bias upward the
estimate of the effect of the pension on weight for height.
A second problem is the possibility that the pension program led to changes
in the composition of the household. Endogenous household composition could
create a positive correlation between unobserved household characteristics and
the presence of an eligible member. A difference between the coefficient of a
woman's pension and that of a man's could then be obtained even in the absence
of any causal effect of the additional income on nutrition.
III. Effect of the Pension on Height for Age
Some of these problems, which plague any cross-sectional comparison of house-
holds by eligibility status, could be addressed by comparing health status in
households with eligible members with that in households without eligible mem-
bers, before and after the expansion of the pension program. There were no rep-
resentative surveys of African households before the end of apartheid. But we
can take advantage of the fact that height is a stock reflecting accumulated in-
vestment in a child's health and nutrition since birth.
Empirical Specification
In developing economies human growth deficits are caused by two preventable
factors: inadequate food and infections. Genetic factors matter in child height,
but they become more critical in adolescence. The height for age of young chil-
dren depends on accumulated investments over the life of the child (Martorell
and Habicht 1986).
I capture this by writing the height for age of child i at age a as a function of
nutrition since birth: hi(a) = f(N0 . . . , Nai), where hi(a) is the height-for-age
i
z-score attained by child i at age a, and Nsi (for s = 0 to a) is the ratio of the
nutrition and other necessary inputs (primary health care and the like) received
by the child relative to what would be optimal at each age. The function f(.) is
weakly increasing in all its arguments, and f(1, 1 . . . , 1) = 0.
Some properties of the function f(.) are documented in the medical literature.
First, nutrition at a very early age (in the womb and in infancy) has long-lasting
effects on child height and indeed on adult health (Barker 1990; Scrimshaw 1997).
Second, the possibility of catch-up skeletal growth after an episode of low growth
in infancy is limited.14 Most stunting and catch-up occurs between 6 and 24
months of age. Stunting after 24 months of age generally reflects the interaction
of nutrition and infection at earlier ages (Martorell and Habicht 1986).
14. For example, a study in Jamaica shows that children's weight for height recovers quickly from
episodes of acute malnutrition, but that once normal weight for height is achieved, the body stops ac-
cumulating nutrients that would allow faster skeletal growth (Ashworth 1969).
Duflo 13
Given this, if households eligible for pensions have worse characteristics than
noneligible households, older children would be smaller in eligible households.
But if the expansion of the pension program led to better nutrition, children
measured when they were younger would have been well nourished for a larger
fraction of their lives. Therefore, the younger the children are, the smaller their
relative disadvantage should be in eligible households. If the pension program
led to a substantial improvement in nutrition (as suggested by the previous sec-
tion), the youngest children in eligible households should even be taller for their
age than children of the same age in noneligible households.
The basic idea of the identification strategy is thus to compare the difference
in height between children in eligible and those in noneligible households among
children exposed to the program for a fraction of their lives to the same differ-
ence among children exposed all their lives. Figure 1 illustrates this identifica-
tion strategy. I run a nonparametric regression of height for age on age in months
for children living with an eligible woman, for those living with an eligible man,
and for those living with no eligible household members (children living with
both an eligible man and an eligible woman are included in both regressions).
The curves have the traditional pattern for height for age in developing coun-
tries (Martorell and Habicht 1986). Height for age declines steeply in the first
two years of life and then stabilizes.
The relative position of the curves is of interest. Older children living with an
eligible woman are smaller than those with no eligible member in their house-
hold. The relative advantage of noneligible children increased for children born
between June and December 1990 (noneligible children in that age group ap-
peared to be taller for their age). Starting in January 1991 (when the pension
program began to expand), the difference becomes smaller, and children born
by the end of 1991 are taller if they live with an eligible woman. Duflo (2000a)
presents more nonparametric evidence of the effect of the program (without
distinguishing by the recipient's gender), showing that the difference in its effect
for eligible and noneligible girls has a significant positive slope after January 1991.
This catch-up is not apparent for young children living with an eligible grandfa-
ther,15 even though the height for age of the older children living with an eligible
grandfather is very similar to that of children living with an eligible grandmother.
This discussion suggests the following formulation for comparing the effect
of a woman's pension eligibility on height for age with that of a man's:
4
(3) hijk = pf(YOUNG * Ef) + pm(YOUNG * Em) + bfEf + bmEm + gj1( l=k)
4 l=1
+ Xijkd + 1( l=k)* Xijklj + eikk
l=1
where hijk is the height-for-age z-score of a child born in cohort k in family j, and
the notation is otherwise as before. Children born in January 1992 or later, after
the full expansion of the pension program, form the most exposed group
15. The catch-up at the extreme right of figure 1 is due to one outlier.
14 the world bank economic review, vol. 17, no. 1
Figure 1. Height for Age of Children Living with Eligible Women, Eligible
Men, No Eligible Member
4
(YOUNG). The last two terms (Xijk and 1( * Xijk) are family background
j=1 l=k)
variables (those discussed in the previous section, plus a control variable for the
presence of a household member over age 50) and family background vari-
ables interacted with cohort dummy variables. Equation 3 is estimated using
ordinary least squares (ols), and standard errors are adjusted to take into ac-
count the correlation of error terms between children in the same families as
well as heteroscedasticity.
Because even older children were exposed to the program for a fraction of
their lives, the difference-in-differences estimate is a downward-biased estimate
of the effect of the eligibility (and of the difference between the effects of men's
and women's eligibility) unless nutrition at a very early age is the only signifi-
cant determinant of height by age five.
Results
The results from estimating equation 3 are presented in table 5. Column 1 esti-
mates the effect of pension eligibility. For girls, living with an eligible household
member is associated with an increase of 0.68 standard deviation in height for
age. The uninteracted effect of eligibility is negative but substantially smaller than
the interaction (�0.17) and insignificant after the control variables are introduced.
This result is reassuring, because it reduces the likelihood that eligible and non-
eligible households are subject to different shocks (such as different programs).
Duflo 15
Table 5. ols and 2sls Regressions of the Effect of Pension Eligibility,
Presence of an Old Grandparent, and Pension Receipt
Treatment variable
Old Receives
Eligibility Eligibility grandparent pension
ols ols ols 2sls
(1) (2) (3) (4)
Girls
Eligible household � YOUNG 0.68*
(0.37)
Woman treatment variable � YOUNG 0.71* 0.40 1.16*
(0.34) (0.27) (0.56)
Man treatment variable � YOUNG 0.097 �0.12 �0.071
(0.57) (0.35) (0.95)
Eligible household �0.17
(0.16)
Woman pension variable �0.15 �0.039 �0.15
(0.17) (0.13) (0.17)
Man pension variable �0.11 0.027 �0.11
(0.24) (0.15) (0.24)
Observations 1533 1533 1533 1533
Boys
Eligible household � YOUNG 0.11
(0.31)
Woman pension variable � YOUNG 0.18 0.026 0.28
(0.32) (0.27) (0.47)
Man pension variable � YOUNG �0.30 0.18 �0.47
(0.32) (0.30) (0.71)
Eligible household �0.15
(0.15)
Woman pension variable �0.14 �0.084 �0.15
(0.32) (0.69) (0.17)
Man pension variable �0.073 �0.011 �0.057
(0.21) (0.14) (0.21)
Observations 1627 1627 1627 1627
Control variables
Age dummy variablesa Yes Yes Yes Yes
Family background variablesb Yes Yes Yes Yes
Family background variables � Yes Yes Yes Yes
age dummy variables
*Significant at the 10 percent level.
Note: Standard errors (robust to correlation of residuals within households and heteroscedasticity)
are in parentheses.
aDummy variables for whether the child was born in 1991, 1990, or 1989.
bFather's age and education; mother's age and education; rural or metropolitan residence; size of
household; and number of members ages 0�5, 6�14, 15�24, 25�49, and over 50.
Source: Author's calculations.
16 the world bank economic review, vol. 17, no. 1
For boys the effect of eligibility is small (0.11) and insignificant. The uninteracted
effect of eligibility is similar to that on girls' height (�0.15).
Column 2 estimates the effect distinguishing by gender of the household mem-
ber eligible for the pension. Having an eligible woman in the household increases
the height for age of young girls relative to older ones by 0.71 standard devia-
tion (with a standard error of 0.34). Having an eligible man in the household
has a small and negative effect. For boys the coefficient of the interaction be-
tween a woman's eligibility and a dummy variable for being young is only 0.18
(and insignificant) and that of the interaction between a man's eligibility and a
dummy variable for being young is negative and insignificant.
As in the weight-for-height specification, I estimate the effect of the pension
implied by these coefficients using 2sls:
(4) hijk = af(YOUNG * PENSf) + am(YOUNG * PENSm) + bfEf + bmEm
4 4
+ gj1( j=k)+ Xifkd + 1( k=j)* Xifklj + eifk
j=1 j=1
where PENSf is a dummy variable equal to one if a woman receives a pension,
and PENSm a dummy variable equal to one if a man receives a pension. The
interactions YOUNG * PENSf and YOUNG * PENSm are endogenous, and they
are instrumented using the interactions (YOUNG * Ef) and (YOUNG * Em).16
The results suggest that pensions received by women led to an increase of at
least 1.16 standard deviations in the height of girls and to a much smaller (and
insignificant) effect (0.28 standard deviation) on the height of boys (table 5,
column 4). Pensions received by men appear to have had no effect on the height
of boys or girls. These results are strikingly similar to those for weight for height.
In table 6 I consider separately the effect of pension eligibility of the mother's
mother, the father's mother, the mother's father, and the father's father. Again,
the eligibility of the mother's mother had the strongest effect on the nutritional
status of girls.
It is reassuring to see that the two outcome measures (weight for height and height
for age) and the two strategies lead to the same results. Nevertheless, these results
could be tainted by failures of the identification assumption, which I now consider.
Controlling for Endogenous Household Formation
and for Other Programs
The comparison between older and younger children in eligible and noneligible
households helps address some of the earlier concerns. In particular, any differ-
ence between eligible and noneligible households that affects older and younger
16. An alternative specification would be to control for uninteracted pension receipt, instrumented
by the uninteracted eligibility variable. The reduced form would be identical, and the coefficients of the
interaction YOUNG * PENS similar. Because the main effect of PENS cannot be interpreted, this is the
preferred specification.
Duflo 17
Table 6. Effect of Pension Eligibility on Height for Age by
Gender of the Intermediate Generation: ols Regressions
Girls Boys
Mother's mother eligible � YOUNG 0.94 0.23
(0.56) (0.51)
Father's mother eligible � YOUNG 0.76 �0.34
(0.53) (0.54)
Mother's father eligible � YOUNG �0.69 �0.70
(0.65) (0.82)
Father's father eligible � YOUNG 0.33 0.36
(0.69) (0.81)
Observations 1457 1552
Control variables
Family background variablesa Yes Yes
Age dummy variablesb Yes Yes
Eligibility variablesc Yes Yes
Family background variables � Yes Yes
age dummy variables
Note: Standard errors (robust to correlation of residuals within households
and heteroscedasticity) are in parentheses.
aFather's age and education; mother's age and education; rural or metro resi-
dence; size of household; and number of members ages 0�5, 6�14, 15�24, 25�49,
and over 50.
bDummy variables for whether the child was born in 1991, 1990, or 1989.
cMother's mother eligible, father's mother eligible, mother's father eligible,
and father's father eligible (the first stage is in table A-1).
Source: Author's calculations.
children in a similar way is controlled for. Even so, there could still be age-specific
differences in nutritional status across households.
Endogenous Household Formation. As discussed, household composi-
tion may have changed as a result of the program, and this could invalidate
the proposed identification strategy if parents became more (or less) likely to
send their children to live with their grandparents or to have the grandparents
live with them. If this did not depend on the age of the child, this would not
invalidate the spirit of the strategy. But it is conceivable that the correlation
between child health and household composition for young children differs from
that for older children.
To address this problem, I use an alternative variable that is correlated with
the presence of an eligible member in the household but is not affected by house-
hold decisions. This is a dummy variable indicating whether the child has at
least one grandparent who is alive and eligible or likely to be eligible. The
dummy variable takes the value of one if there is an eligible person in the house-
hold or if one of the following is true: the mother (the father) of the child is
older than age 34 and her (his) mother is alive, or the mother (father) of the
18 the world bank economic review, vol. 17, no. 1
child is older than age 32 and her (his) father is alive.17 Among children who
have a living old grandparent, 46 percent live with a pension recipient. This
variable is therefore still strongly correlated with pension receipt. Using the
variable as an alternative instrument for pension receipt would be inappropri-
ate, because it will also capture possible changes in transfers by (or to) a
noncoresident grandparent.18
Estimates of a difference-in-differences specification that is similar to the re-
duced form equation 3 but uses the indicator for whether the child has a living
elderly grandparent rather than eligibility status are presented in table 5 (col-
umn 3). For girls, having a living old grandmother has a positive (but significant
at only 15 percent) effect, though it is smaller than the estimated effect of a
woman's eligibility. This result is expected, because the probability of receiving
the pension is higher for those living with an eligible household member than
for those with a living elderly grandparent, and interhousehold transfers may
not fully compensate for this difference. The effect of having a living elderly
grandfather is small and insignificant. For boys, having a living old grandfather
or grandmother has no effect.
Using this alternative variable reduces the precision of the results, but it does
not change the conclusions. Thus, it confirms that the previous results are not
likely to be an artifact of endogenous household formation.
Unobserved Characteristics of Government Programs. Could the results
be explained by age-specific differences in nutritional status between children
living in households with an eligible woman and children living in other
households?
Younger girls are taller in households with an eligible woman relative to girls
of the same age in noneligible households. When I estimate the effect of eligibil-
ity on weight for height (equation 1) separately for younger and older girls, the
coefficients of a woman's pension eligibility are 0.71 and 0.56. Both numbers
differ significantly from zero but not from each other. These results are there-
fore unlikely to be driven by the fact that all children receive worse nutrition
when they live with an eligible grandmother but that the effects are stronger for
older children. The similarity of the effects on the weight for height of younger
and older children suggests that the results are not due to a program targeting
younger children in eligible households.
Similar results would still be obtained if children (older and younger) living
with eligible grandmothers were also more likely to be targeted by other nutri-
17. I determined the cutoffs of 32 and 34 years by using the information on extended families in my
sample. Women whose observed child is older than 34 and men whose observed child is older than 32
have a 60 percent probability of being eligible for the pension. Results are not sensitive to the choice of
cutoff. If a parent is not in the household, the survey does not indicate his or her age or whether his or
her parents are alive. So some children may have a living old grandparent not identified in the data.
18. Jensen (1998) shows that intergenerational transfers changed in response to the program: trans-
fers from children to parents were reduced when parents became eligible.
Duflo 19
tion or government health programs. South Africa had two nutritional programs
in place in the early 1990s (Budlender 2000). The Protein Energy Malnutrition
Scheme, in place since the 1960s, subsidized the purchase of powdered skim milk
for distribution to malnourished preschool children. The program was modi-
fied and expanded in mid-1993 (its annual budget was increased to R40 mil-
lion, eight times its former budget, and its target group was broadened). Because
the sample excludes children born later than July 1993, it would not be affected
by this expansion. The National Development Program (later renamed the Na-
tional Nutrition and Social Development Program) was introduced in 1990�91
to compensate for the planned introduction of a value-added tax on basic food-
stuffs, with an annual budget of R400 million. This food distribution program
(not particularly focused on children) was implemented at the local level through
nongovernmental and community-based organizations. Because of the program's
decentralized implementation, it is impossible to document which households
benefited more. But because those eligible for pensions are generally poorer, they
are more likely to have been in the target groups.
Three pieces of evidence suggest that these programs do not account for the
results. First, the characteristics of households with eligible women and those
with eligible men are very similar. Thus, it appears unlikely that a program
would have disproportionately targeted children living with their grandmothers
rather than all children living in extended families. Of course, grandmothers
might be more likely than grandfathers to take advantage of these programs.
But the interaction between the child's age and the presence of a woman over
age 50 (but not eligible) has the same coefficient as that between the child's
age and the presence of man over age 50, providing no support for this alter-
native explanation. Second, the regressions control for a range of observed
household variables interacted with age dummy variables and should there-
fore capture the effect of any program targeted according to these variables.
When these variables are controlled for, older children are not significantly
smaller in eligible households, suggesting that they are unlikely to have been
targeted by other programs. Third, the point estimates obtained with the al-
ternative instruments (grandmother or grandfather alive and old) are similar
to those obtained using eligibility variables as instruments. The characteristics
of households in which children have a living grandmother are similar to those
in which children do not have one. Thus it is unlikely that they would have
been subject to different programs.
IV. Interpretation
The evidence appears to indicate that pensions received by women had a large
effect on child nutrition, whereas pensions received by men did not. But several
interpretations of this evidence are possible.
One interpretation is that the same resources are spent differently depending
on whether they are received by a woman or by a man. This interpretation would
20 the world bank economic review, vol. 17, no. 1
have important implications for public policy. In particular, it suggests that if
improving children's nutrition is a policy objective, targeting public transfers to
women rather than to men might be preferable.
A second interpretation is that, in terms of permanent income, a rand of pen-
sion received by a man represents much less than a rand of pension received by
a woman because men are expected to receive the pension for a shorter time.
This difference could lead to different effects from men's and women's pensions
if households have an ability to smooth consumption over time through savings
or borrowing.
To help discriminate between these two interpretations, it is useful to look at
the disposition of men's and women's pension income. If the household is a
unitary entity and if a man's pension income is not spent on child health because
it is akin to transitory income, the propensity to save out of a man's pension
income should be much larger than the propensity to save out of a woman's
pension income (and nonpension income). To examine this possibility, I estimate
the following equation:
(5) Sh = afyhf + amyhm + azh + Xhb + eh
where Sh is the total savings of the household (defined as total income minus
expenditures), yhf is pension income received by a woman, yhm is pension income
received by a man, zh is nonpension income, and Xh is a set of control variables.
This specification extends the Case and Deaton (1998) formulation to take into
account differences in the disposition of income received by men and women.
The emphasis here is on the comparison between af and am. The equation is es-
timated using both ols and 2sls. The instruments in the 2sls equations (for
yhf, yhm, and zh) are the indicators for the presence of an eligible man and an
eligible woman and the instruments used to correct for measurement errors in
nonpension income (see notes to table 7).
The point estimates suggest that the propensity to save out of a man's pen-
sion income is lower than the propensity to save out of a woman's pension in-
come, although the difference is not significant (table 7).19 This result indicates
that the differences in the effects of women's and men's pension income on child
height are unlikely to be due to the differences in their life cycles.
The argument could then be reversed. If child nutrition is an investment,
grandmothers' expectations of a longer life would lead them to invest more in
their grandchildren, because they are more likely to reap the benefits of this
investment. This could explain the fact that the elasticity of child nutrition with
respect to the grandmother's income is larger than that with respect to the
grandfather's. In this case the differences in the effect of pension income on
child nutrition have nothing to do with the gender of the recipient per se; in-
stead, the differences result from the fact that, in this particular program, female
19. The very large estimated propensity to save out of nonpension income should not be taken at
face value. It reflects mismeasurement of income and consumption, common in this type of survey.
Duflo 21
Table 7. Propensity to Save out of Pension Income
and Nonpension Income: ols and 2sls Regressions
Savings
Variable ols 2sls
Woman's pension income 0.99* 0.82*
(0.093) (0.16)
Man's pension income 0.78* 0.53*
(0.13) (0.22)
Nonpension income 0.53* 0.50*
(0.017) (0.041)
*Significant at the 5 percent level.
Note: Standard errors are in parentheses. Instruments are dummy
variables for household head is employed; household head holds a regu-
lar job, a casual wage job, a job in agriculture; sector of the job; type
of employer (central or local government, private firm, other); type of
pay (weekly, fortnightly, monthly); woman eligible; and man eligible.
Source: Author's calculations.
recipients are younger and live longer. However, when I reestimate the rela-
tionships of child height for age and weight for height in a sample that includes
only eligible men living with an eligible woman, the results (not presented here)
are unchanged.
Grandmothers are likely to have a stronger incentive than grandfathers to
invest in children because they will benefit from them for a longer time. But the
fact that this difference in preferences results in a difference in outcomes shows
that individual preferences and bargaining power matter for how expenditures
are allocated. These results thus provide new evidence that households do not
function as a unitary entity (Chiappori 1988, 1992; Browning and Chiappori
1998), evidence untainted by the empirical problems (such as assortative matching
and endogeneity of income) affecting earlier studies.20 This in turn suggests that
the identity of the recipient is an important parameter in the design of a public
transfer program, even though it cannot be inferred from these results that grand-
mothers have a stronger inherent preference for children.
V. Conclusion
The expansion of the old-age pension program in South Africa led to an improve-
ment in the health and nutrition of girls, reflected in the weight for height of all
girls and the height for age of the youngest girls. It had no discernible effect on
boys. The effect is entirely due to pensions received by women.
Pensions received by women improved the height-for-age z-scores of younger
girls by at least 1.16 standard deviations, and the weight-for-height z-scores of
20. In an earlier version of this article I argue that to convincingly reject the unitary model of the
household, one needs exogenous permanent shocks to the nonlabor income of both household mem-
bers, occurring after marriage, which is exactly what the pension program provides (Duflo 2000b).
22 the world bank economic review, vol. 17, no. 1
girls by 1.19 standard deviations. South African girls are on average 1.20 stan-
dard deviations smaller than U.S. children, so the estimates suggest that pensions
received by women helped girls bridge the entire gap in stature with U.S. children.
The pension represented a large income transfer, so a finding of large effects
on child nutritional status perhaps appears unsurprising. Still, the fact that pen-
sions received by women led to a sizable increase in the height of girls shows
that cash transfers can have an important effect on child nutritional status. There
is almost no evidence of this kind for developing economies,21 but the available
evidence for the United States suggests that cash transfers do not have substan-
tial effects on child welfare (Mayer 1997; Currie 1995).22 One would expect these
effects to be larger in developing economies, where households are more likely
to face credit constraints; this article confirms that intuition.
Of course, the article's findings cannot be easily generalized to other develop-
ing economies. The old-age pension program was on such a large scale that it
could not be replicated outside the particular context of the postapartheid era in
South Africa. Moreover, there could be nonlinearity in the effect of a cash trans-
fer, making it difficult to infer what the effect would have been had it been imple-
mented on a different scale.
Thus, the most important finding may be that this large cash transfer had no
effect if it was received by a man. This suggests that the efficiency of transfer
programs may vary depending on how they are administered. In South Africa
the program is naturally biased toward women, both because women can claim
the pension earlier (at age 60, compared with age 65 for men) and because women
tend to live longer. Without this feature the program would have a smaller effect
on the nutrition of young children.
The distinction between men and women does not accord with the South Afri-
can constitution, and there is some pressure to remove it. But the effectiveness of
the pension program as a tool for transferring resources to young children would
argue for increasing the bias toward women. Even so, pensions received by men
could affect other dimensions of investment in children's human capital not mea-
sured here (such as education), so this implication needs to be carefully assessed.
Future work should investigate whether the difference between women and men--
and between girls and boys--is also found for other outcomes. Moreover, it is
important to understand the reason for the larger effect on girls. Is it because they
were lagging further behind?23 Is the effect specific to income received by grand-
mothers? If so, why do grandmothers prefer girls?
21. An exception is a study by Carvalho (2000) showing that an expansion of the old-age pension
program in Brazil led to an increase in educational attainment among girls and to a decrease in child
labor among boys.
22. Shea (1997) studies whether outcomes for children (education and subsequent labor earnings)
in the United States are correlated with their father's union status, job loss, or industry of employment
and finds no effect except among the poorest families.
23. The available evidence does not allow an answer to this question. z-Scores cannot be easily inter-
preted, because the reference population consists of well-nourished U.S. children. Nor can they be com-
pared across genders, because the standardization may distort the data in different ways for boys and girls.
Duflo 23
Appendix
Table A-1. Effect of Pension Eligibility on Pension Receipt: First-Stage
Regressions
Woman Man
Woman Man receives receives
receives receives pension pension
pension pension � YOUNG � YOUNG
(1) (2) (3) (4)
Girls
Woman eligible 0.51* 0.025
(0.057) (0.03)
Man eligible 0.077 0.41*
(0.052) (0.064)
Woman eligible � YOUNG 0.62* 0.017
(0.062) (0.042)
Man eligible � YOUNG 0.12 0.60*
(0.10) (0.094)
Observations 1533 1533 1533 1533
Boys
Woman eligible 0.55* 0.021
(0.052) (0.036)
Man eligible 0.028 0.39*
(0.054) (0.067)
Woman eligible � YOUNG 0.70* 0.025
(0.050) (0.045)
Man eligible � YOUNG �0.071 0.59*
(0.059) (0.069)
Observations 1627 1627 1627 1627
Control variables
Presence of older membersa Yes Yes No No
Family background variablesb Yes Yes Yes Yes
Family background variables � Yes Yes Yes Yes
age dummy variables
Age dummy variablesc Yes Yes Yes Yes
*Significant at the 5 percent level
Note: Standard errors (robust to correlation of residuals within households and heteroscedasticity)
are in parentheses.
aDummy variables for whether there is a woman over age 50, a man over age 50, a woman over age
56, a man over age 56, a man over age 61.
bFather's age and education; mother's age and education; or metropolitan residence; size of house-
hold; and number of members ages 0�5, 6�14, 15�24, 25�49, and over 50.
cDummy variables for whether the child was born in 1991, 1990, or 1989.
Source: Author's calculations.
References
Ashworth, Ann. 1969. "Growth Rates in Children Recovering from Protein-Calorie
Malnutrition." British Journal of Nutrition 23:835�45.
Balazs, R., T. Jordan, P. D. Lewis, and A. J. Patel. 1986. "Undernutrition and Brain
24 the world bank economic review, vol. 17, no. 1
Development." In F. Falkner and J. M. Tanner, eds., Human Growth: A Comprehen-
sive Treatise, vol. 3, 2d ed. New York: Plenum.
Barker, D. J. P. 1990. "The Fetal and Infant Origins of Adult Disease." British Medical
Journal 301(6761):1111.
Bertrand, Marianne, Douglas Miller, and Sendhil Mullainathan. 1999. "Public Policy
and Extended Families: Evidence from South Africa." Massachusetts Institute of Tech-
nology, Department of Economics, Cambridge, Mass.; and Princeton University, De-
partment of Economics, Princeton, N.J.
Browning, Martin, and Pierre-Andr� Chiappori. 1998. "Efficient Intrahousehold
Allocations: A General Characterization and Empirical Tests." Econometrica
66(6):1241�78.
Budlender, Debbie. 2000. "Human Development." In Julian May, ed., Poverty and In-
equality in South Africa: Meeting the Challenge. Johannesburg: Zed.
Carvalho, Irineu. 2000. "Household Income as a Determinant of Child Labor and School
Enrollment in Brazil: Evidence from a Social Security Reform." Massachusetts Insti-
tute of Technology, Department of Economics, Cambridge, Mass.
Case, Anne, and Angus Deaton. 1998. "Large Cash Transfers to the Elderly in South
Africa." Economic Journal 108(450):1330�61.
Chiappori, Pierre-Andr�. 1988. "Rational Household Labor Supply." Econometrica
56(1):63�90.
------. 1992. "Collective Labor Supply and Welfare." Journal of Political Economy
100(3):437�67.
Currie, Janet. 1995. Welfare and the Well-Being of Children; Fundamentals of Pure and
Applied Economics 59. Chur, Switzerland: Harwood Academic.
Dasgupta, Partha. 1993. Inquiry into Well-Being and Destitution. Oxford: Clarendon.
Duflo, Esther. 2000a. "Child Health and Household Resources in South Africa: Evidence
from the Old Age Pension Program." American Economic Review 90(2):393�98.
------. 2000b. "Grandmothers and Granddaughters: Old Age Pension and Intra-House-
hold Allocation in South Africa." nber Working Paper 8061. National Bureau of Eco-
nomic Research, Cambridge, Mass.
Edmonds, Eric, Kristin Mammen, and Douglas Miller. 2001. "Rearranging the Family?
Household Composition Responses to Large Pension Receipts." Dartmouth College,
Department of Economics, Hanover, N.H.
Gertler, Paul, and Simone Boyce. 2001. "An Experiment in Incentive-Based Welfare: The
Impact of Progresa on Health in Mexico." Haas School of Business, University of
California at Berkeley.
Jensen, Robert. 1998. "Three Essays on Microeconomics and Social Policy in South
Africa." Ph.D. diss., Princeton University, Department of Economics. Princeton, N.J.
Le Roux, Pieter. 1995. "Poverty, Social Policies and the Reconstruction and Develop-
ment Programme." Working Paper. Institute for Theological and Interdisciplinary
Research, Cape Town, South Africa.
Lund, Frances. 1993. "State Social Benefits in South Africa." International Social Secu-
rity Review 46(1):5�25.
Lundberg, Shelly J., Robert A. Pollak, and Terence J. Wales. 1996. "Do Husbands and
Wives Pool Their Resources? Evidence from the United Kingdom Child Benefit." Jour-
nal of Human Resources 32(4):463�80.
Duflo 25
Martorell, R., and J. P. Habicht. 1986. "Growth in Early Childhood in Developing Coun-
tries." In F. Falkner and J. M. Tanner, eds., Human Growth: A Comprehensive Trea-
tise, vol. 3, 2d ed. New York: Plenum.
Mayer, Susan E. 1997. What Money Can't Buy: Family Income and Children's Life
Chances. Cambridge, Mass.: Harvard University Press.
Miguel, Edward, and Michael Kremer. 2001. "Worms: Education and Health Externali-
ties in Kenya." nber Working Paper 8481. National Bureau of Economic Research,
Cambridge, Mass.
Schultz, T. Paul. 1999. "Productive Benefits of Improving Health: Evidence from Low-
Income Countries." Yale University, Department of Economics, New Haven, Conn.
------. 2000. "Final Report: The Impact of Progresa on School Enrollments." Yale
University, Department of Economics, New Haven, CT; and International Food Policy
Research Institute, Washington, D.C.
Scrimshaw, Nevin S. 1997. "The Relation between Fetal Malnutrition and Chronic Dis-
ease in Later Life." British Medical Journal 315:825�26.
Shea, John. 1997. "Does Parents' Money Matter?" nber Working Paper 6026. National
Bureau of Economic Research, Cambridge, Mass.
Strauss, John, and Duncan Thomas. 1998. "Health, Nutrition, and Economic Develop-
ment." Journal of Economic Literature 36(2):766�817.
Thomas, Duncan. 1990. "Intrahousehold Resource Allocation: An Inferential Approach."
Journal of Human Resources 25(4):635�64.
------. 1994. "Like Father, Like Son, Like Mother, Like Daughter: Parental Education
and Child Health." Journal of Human Resources 29:950�88.
Van der Berg, Servaas. 1994. "Issues in South African Social Security." Office memoran-
dum. World Bank, Washington, D.C.
who (World Health Organization) Working Group. 1986. "Use and Interpretation of
Anthropometric Indicators of Nutritional Status." Bulletin of the World Health Organi-
zation 64(6):929�41.
the world bank economic review, vol. 17, no. 1 27�50
Public Policy and Extended Families:
Evidence from Pensions in South Africa
Marianne Bertrand, Sendhil Mullainathan, and Douglas Miller
How are resources allocated within extended families in developing economies? This
question is investigated using a unique social experiment: the South African pension
program. Under that program the elderly receive a cash transfer equal to roughly twice
the per capita income of Africans in South Africa. The study examines how this trans-
fer affects the labor supply of prime-age individuals living with these elderly in extended
families. It finds a sharp drop in the working hours of prime-age individuals in these
households when women turn 60 years old or men turn 65, the ages at which they
become eligible for pensions. It also finds that the drop in labor supply is much larger
when the pensioner is a woman, suggesting an imperfect pooling of resources. The allo-
cation of resources among prime-age individuals depends strongly on their absolute
age and gender as well as on their relative age. The oldest son in the household reduces
his working hours more than any other prime-age household member.
In many developing economies large extended families often live together. Shared
housing may suggest the sharing of other resources, most notably money. If such
resource sharing is prevalent, social policies may produce unexpected outcomes.
A transfer targeted to one demographic group may eventually find itself in the
pockets of relatives living in the same house. Who in the end benefits from the
transfer will depend on the sharing rules within the household.
To understand how resources are transferred in extended families,1 this study
investigates South Africa's unusual old-age pension program. The program grants
Marianne Bertrand is Associate Professor of Economics, University of Chicago Graduate School of
Business, Center for Economic and Policy Research, and National Bureau of Economic Research. Her
e-mail address is marianne.bertrand@gsb.uchicago.edu. Sendhil Mullainathan is Associate Professor
of Economics, Massachusetts Institute of Technology and National Bureau of Economic Research. E-mail:
mullain@mit.edu. Douglas Miller is Assistant Professor of Economics, University of California, Davis.
His e-mail address is dlmiller@uc.davis.edu. The authors are grateful to two anonymous referees, Abhijit
Banerjee, Anne Case, Angus Deaton, Esther Duflo, Jon Gruber, Michael Kremer, Jonathan Morduch,
and Jim Poterba for many helpful comments. They have also benefited from feedback from seminar
participants at the MIT Public Finance Lunch, Princeton Development Workshop, Harvard-MIT Devel-
opment Seminar, and the National Bureau of Economic Research Summer Institute 2000. Miller ac-
knowledges financial support from the National Science Foundation's Graduate Fellowship Program.
1. A large body literature has examined resource transfers in the close family (husband and wife or
parent and young child). Lundberg and Pollak (1996) provide a survey. In the close family, one can rea-
sonably assume that resource transfers take place, for example, between parents and young children. The
DOI: 10.1093/wber/lhg014
� 2003 The International Bank for Reconstruction and Development / THE WORLD BANK
27
28 the world bank economic review, vol. 17, no. 1
large lump-sum cash transfers, roughly twice the average per capita income in
African households, to eligible women over the age of 60 and men over the age
of 65.2 The magnitude of the transfer makes it a useful experiment, permitting
the tracking of the flow of money more cleanly than would more marginal
changes. Does the pension money eventually reach family members other than
the pensioners? If so, how much of the cash is transferred, and which family
members receive most of it?
These questions are addressed through an examination of the labor supply of
relatives living with pensioners. This approach has two advantages. First, typical
household survey data do not allow direct measuring of the transfers to each fam-
ily member. The survey data used in this study are no exception. Expenditure data
measure consumption at the household, not the individual level. Only a few con-
sumption items are exclusive enough that they can be matched to a specific gen-
der or age group.3 Leisure time, however, is a good that can easily be assigned
(Chiappori 1992). Labor supply data can be used to infer (at least partly) how the
pension money is allocated among the prime-age individuals in a household.
Second, a labor supply response would most clearly underline the unexpected
outcomes caused by family redistribution. Because the social pension targets a
group that by and large is already out of the labor force and conditions mainly
on an unalterable variable, age, it might be expected to have little effect on labor
supply.4 On the other hand, if intrahousehold redistribution occurs, aggregate
labor supply may fall as the prime-age individuals who live with pensioners re-
duce their hours of work. The size of the effect will depend on the direction and
strength of redistribution flows inside households.
Anecdotal evidence and newspaper articles hint that the pension may well have
affected relatives' labor supply. One article mentions that "the impact of pen-
sions on communities with a high rate of unemployment was huge, as multi-
generation households formed a constellation around the person receiving the
pension" (Ngoro 1998). Another describes a pensioner's "five children, who also
live with him in his two-bedroom flat, contribute to the family income when they
can find work. But none has a full-time job" (Caelers 1998). Of course, such
rare evidence on resource transfers in the extended family comes from the United States (Altonji and
others 1992) and suggests that resource transfers are not large in that case. It is an open question whether
such a finding generalizes to developing economies, where the extended family often lives under a com-
mon roof.
2. The survey data used for this study classified people into four different racial groups: white, col-
ored, Indian, and African. This study looks only at African households. In theory, the transfer program
is means-tested, but this has little effect in practice for Africans whose income is quite low relative to
the test.
3. Subramanian and Deaton (1991) use expenditures on adult goods, such as alcohol and tobacco,
to study discrimination based on children's gender. Browning and others (1994) use expenditures on
men's and women's clothing to study sharing rules between couples.
4. A labor supply effect might arise because the pension increases the expected future income of the
young. But this would affect all the young equally, not merely the relatives of pensioners, as the results
suggest.
Bertrand, Mullainathan, and Miller 29
stories do not constitute causal evidence, but they provide a backdrop for the
statistical work presented here.
The study uses the sharp rise in household income when an elderly member
crosses the pension age threshold to identify the pension's effect. The results
suggest that the pension dramatically reduces the labor supply of the prime-age
members of the household. Both hours worked and the work or not-work margin
are affected. A clear discontinuity appears exactly at the age-eligibility frontier,
with labor supply by the household dropping when a woman in the household
reaches age 60 or a man reaches age 65. Roughly speaking, the age of the elderly
does not seem to affect labor supply except at these discontinuous points.
Absolute age, relative age, and gender are important determinants of resource
flows in that they affect the strength of the labor supply response. Holding fam-
ily composition constant, the study finds that the marginal rand of pension in-
come going to a female pensioner reduces labor supply more than the marginal
rand of pension income going to a male pensioner. This gender impact on the
flow of resources suggests that common-preference models of the family, which
view the family as maximizing one common utility function, and for which the
source of the pension income should not matter, may not fit these extended fami-
lies very well.
The study also finds that prime-age women reduce their labor supply less than
prime-age men for each marginal rand of pension money received by the elderly.
Also, working hours drop more as the age of the prime-age family member in-
creases. Finally, after controlling for the differential effect of the pension by gender
and age, the study finds that the oldest prime-age male in a household reduces
his labor supply more than do other prime-age household members.
In summary, although the South African pension program was introduced as
a way to improve living standards among elderly people who do not have access
to a private pension, the results show that intrahousehold redistribution sub-
stantially reduces the size of the transfer to that demographic group. At least
part of the pension money ends up with a group that was not originally targeted:
prime-age individuals that live with the pensioners
I. The Old-Age Pension Program
The social pension program in South Africa, which dates to the 1920s, was his-
torically intended for white South Africans only.5 Disintegration of the apart-
heid regime in the late 1980s and early 1990s led to pressures for more racial
parity in pension eligibility and benefits. Major reforms of the pension program
for African households took place after 1992, with the introduction of superior
technologies in the pension delivery system (in part to improve access to remote
5. Additional information about the historical background, institutional features, and practical
implementation of this program can be found in Lund (1992), Van der Berg (1994), and Case and Deaton
(1998).
30 the world bank economic review, vol. 17, no. 1
areas) and the equalization of both the means-test and the pension benefit levels
across racial groups.
Eligibility for pension receipt is determined primarily by age: only women over
the age of 60 and men over the age of 65 are eligible. In practice, though, some
local authorities have been equalizing the pension eligibility age for men and women.
Hence, a nontrivial share of men between 60 and 65 years of age report receiving
a pension. (This fact is exploited later in the analysis of the effect of the pension.)
The state social pension is means tested, with the result that most whites are
excluded from the pension whereas most Africans are entitled to the maximum
benefits. Case and Deaton (1998) show that 14 percent of white women and
7 percent of white men report receiving the pension, compared with 80 percent
of African women and 77 percent of African men.6
The South African social pension is very generous. The maximum benefit in
1993, the year the survey data were collected, was 370 rand a month, or about
half the average African household income and more than twice the median per
capita income among Africans. Such large pension transfers could be expected
to result in intrahousehold redistribution that leads to significant behavioral
responses, such as a reduced willingness to participate in the labor force among
family members not originally targeted by policymakers.
II. Data and Summary Statistics
The primary data set used in this article is the Integrated Household Survey of
South Africa. This survey is the result of a cooperation between the World Bank
and the South African Development Research Unit at the University of Cape
Town.7 The survey, a random sample of 9000 households, was conducted dur-
ing the second half of 1993. Means-testing for the pension is such that only a
small share of elderly white women and white men report receiving any pension
transfer, and the participation rate for colored and Indian South Africans, though
higher, is well below African rates (Case and Deaton 1998). Moreover, the preva-
lence of multigeneration households is much larger among Africans than among
the other racial groups (Ardington and Lund 1994).
To keep the focus on extended families, the study was restricted to three-
generation households (a household containing at least a child, a parent, and a
grandparent). This restriction also reduces the heterogeneity in the sample. With-
out it, pension-ineligible households could also include individuals living away
from their elders. Because such individuals would clearly be different from those
living with their elders, this could introduce a selection bias. The restriction to
6. The means- testing formula does not take into account income from family members other than
the elderly (Case and Deaton 1998). Hence, there are no direct incentives in the program design for
family dissolution or migration.
7. The database used in this article can be downloaded directly from www.worldbank.org/html/
prdph/lsms.
Bertrand, Mullainathan, and Miller 31
three-generation households guarantees that the age of the elderly is the only
source of variation.
The study looks at the labor supply of working-age individuals between 16
and 50 years old (prime age) in these multigeneration households. A conserva-
tive cut-off age of 50 years is used to avoid any effect arising because people
expect to get the pension themselves soon. More than a third of prime-age indi-
viduals in the original sample live in three-generation households, as do a large
proportion of women over age 60 and men over age 65, a fact previously noted
by Case and Deaton (1998).8
The dependent variable in most of the regressions reported here is weekly
working hours for prime-age individuals. For each person 16 years old or older,
the survey asks: "How many hours did ______ work last week?" The working
hours question relates to all forms of employment: regular wage employment
(self-employed professionals), casual wage employment, self-employment in
agriculture, and other forms of employment and self-employment. The analysis
also sometimes uses a dummy variable for employment status as a measure of
labor supply. Again, the employment status variable refers to all forms of em-
ployment and not exclusively to regular employment.
The study also briefly documents whether any change in employment status
reflects a change in unemployment or labor force participation status. Individu-
als who report not being currently employed are asked if they have been looking
for work during the previous week. Answers to these two questions are used to
classify people as employed, unemployed, or not in the labor force. Individuals
out of the labor force are then asked why they did not look for work in the pre-
vious week. Individuals out of the labor force who did not look for work be-
cause they thought there were "no jobs or work available" are further classified
as discouraged workers.
Table 1 presents means and standard deviations of the main variables of in-
terest for African individuals between the ages of 16 and 50 who live in three-
generation households. Because identification of the pension impact eventually
relies on the presence or not of age-eligible people in the household, these means
and standard deviations are also presented separately for households with at least
one age-eligible person (woman over 60, man over 65) and households without.
Several interesting facts emerge from table 1. First, only 23 percent of people
in the sample are employed. The employment rate is 26 percent among men and
21 percent among women. Average working hours, 6.3, are also very low. Of
the remaining 77 percent of people who are not employed, 8 percent are unem-
ployed and 21 percent are discouraged, leaving roughly 48 percent out of the
labor force and not discouraged. The low employment rate and high discour-
agement and unemployment rates among prime-age African individuals is a well-
documented characteristic of South African labor markets.
8. As expected, households that contain eligible elderly but that are not three-generation house-
holds are on average much smaller (a little less than four people on average) and older.
32 the world bank economic review, vol. 17, no. 1
Table 1. Descriptive Statistics, 16- to 50-Year-Old Africans in
Three-Generation Households
All Age-eligible Non�age-elegible
households households households
Variable Mean SD Mean SD Mean SD
Age 27.5 9.3 27.5 8.7 27.5 9.9
Employed 0.229 0.420 0.212 0.409 0.246 0.431
Hours worked 6.32 16.37 3.21 12.51 9.45 19.00
Unemployed 0.079 0.270 0.087 0.232 0.071 0.256
Discouraged 0.211 0.408 0.232 0.422 0.191 0.393
4th grade or more 0.754 0.431 0.748 0.434 0.760 0.427
8th grade or more 0.348 0.477 0.338 0.473 0.360 0.480
Matric or more 0.130 0.336 0.128 0.335 0.132 0.338
Household size 8.81 3.62 9.13 3.88 8.50 3.30
Rural 0.683 0.465 0.707 0.455 0.660 0.474
Urban 0.166 0.372 0.152 0.359 0.180 0.384
Metro 0.151 0.358 0.141 0.348 0.161 0.367
Sick 0.065 0.246 0.073 0.261 0.056 0.230
Total income 1325 1833 1318 1246 1333 2272
Pension income 207 275 371 277 42 142
Number of eligible women 0.454 0.526 0.906 0.377 0 0
Number of eligible men 0.169 0.383 0.338 0.485 0 0
Note: Sample is composed of set of African individuals between 16 and 50 years old
that live in a three-generation household. Sample size: all households, 6,326; age-eligible
households, 3,169; non�age-eligible households, 3,157.
Source: All variables are from the World Bank/South African Development Research
Unit survey, August�December 1993.
Second, on background characteristics, the differences between eligible and
ineligible households are small. For example, there are only limited differences
in education or in geographical distribution across rural and urban areas. Age-
eligible households appear a little bigger on average (9.1 versus 8.5).9 One notice-
able difference is that prime-age individuals in eligible households report being
sick more often.10
Third, on employment status and working hours, the difference between the
two household types is dramatic. Raw differences in employment rates are more
than 3 percentage points. The econometric work reported translates these raw
differences into estimates of the effect of the pension.
The analysis shows some other interesting patterns. Pension income in eligible
households accounts for more than a quarter of total household income, dem-
onstrating the generosity of the social pension program. The average eligible
household has 0.9 eligible women and 0.34 eligible men, for a total of 1.24 eli-
9. A similar exercise performed for all prime-age individuals, not only those living in three-genera-
tion households, produces dramatic differences on such variables, underlining the importance of focus-
ing on three-generation households.
10. One might argue that sickness is a luxury good among these African households and that this
might be looked at as an outcome of the social pension.
Bertrand, Mullainathan, and Miller 33
gible members (table 1). Most of the pension income, therefore, comes through
a woman. Many households have more than one pensioner.
III. Basic Results
The first set of regressions compares the labor supply of prime-age individuals who
live with age-eligible elderly relative with those who do not and considers the effect
for both men and women (table 2). Each regression includes, in addition to the
pension variable, a quartic in individual age, a dummy variable for whether the
individual completed eighth grade, 14 province dummy variables, 3 location dummy
variables (rural, urban, and metropolitan area), a female dummy variable, house-
hold size, and number of household members 0�5 years old, 6�15, 16�18, 19�21,
and 22�24 years old.11 For these results and all the results that follow, standard
errors are corrected to allow for correlation in outcomes within household clusters.
Both working hours (columns 1�3) and employment status (columns 4�6) are
used dependent variables (table 2). Basic ordinary least squares (ols) regressions
of labor supply on continuous pension income (columns 1 and 4) show that more
pension income significantly reduces both working hours and employment rates.
The simple ols results, however, are not exploiting only the variation in pen-
sion receipt that comes from the age of the elderly household members. By using
information on actual pension receipt, the estimates may be biased by endog-
enous takeup or eligibility. Takeup rates are high but not complete, and although
the means test is low, some elderly do fail to get the pension. If those who actu-
ally receive the pension are different from those who do not, the ols estimate
will be biased. This possibility is addressed by examining the effect of pension
eligibility (the age-eligibility criterion) rather than actual pension receipt. A similar
negative labor supply response is found for households that have at least one
age-eligible person compared with households that do not (columns 2 and 5).
This eligibility measure cannot easily be transformed into a meaningful economic
measure (such as an elasticity with respect to pension benefits). To ease economic
interpretation, the continuous pension income variable (the amount of pension
benefits received by a household) is instrumented with the number of age-eligible
men and women in that household (columns 3 and 6). The first-stage regressions
associated with columns 3 and 6 (not reported here) show that the numbers of
age-eligible women and men are both very significant determinants of monthly
pension income. The coefficient on number of women over age 60 and number of
men over age 65 are very similar. The null hypothesis that these two coefficients
are the same at standard confidence levels, and hence that the men and women
have similar takeup rates, cannot be rejected. The instrumental variable (iv) coef-
11. The completion of matric (10th grade) is another important determinant of employment and
unemployment probabilities among South African men and women. The results are unaffected if we
use the completion of 10th grade instead of the completion of 8th grade as a control for educational
attainment.
34
the
world
Table 2. Effect of Old-Age Pension Income on Working Hours and Employment Status of 16- to 50-Year-Old Africans
Hours worked Employment status bank
ols ols economic
Pension Pension Pension Pension
uptake eligibility iva uptake eligibility iv
Variableb (1) (2) (3) (4) (5) (6)
review,
Pension income � 1000 �12.32 (1.18) -- �17.07 (1.78) �0.053 (0.022) -- �0.099 (0.035)
Household eligibility -- �6.401 (0.580) -- -- �0.043 (0.011) --
dummy vol.
34 Female �2.552 (0.447) �2.666 (0.452) �2.629 (0.452) �0.068 (0.012) �0.069 (0.012) �0.069 (0.012)
Age �6.526 (3.640) �6.732 (3.709) �6.585 (3.611) �0.394 (0.090) �0.395 (0.090) �0.394 (0.090) 17,
Age2 0.407 (0.187) 0.412 (0.190) 0.404 (0.185) 0.022 (0.055) 0.022 (0.055) 0.021 (0.055)
no.
Age3 �0.010 (0.004) �0.010 (0.004) �0.10 (0.004) �0.0004 (0.0001) �0.0004 (0.0001) �0.0005 (0.0001)
Age4 � 1000 0.082 (0.032) 0.081 (0.032) 0.080 (0.032) 0.0036 (0.0008) 0.0036 (0.0008) 0.0036 (0.0008) 1
8th grade or more 1.485 (0.466) 1.262 (0.468) 1.520 (0.469) 0.064 (0.012) 0.062 (0.012) 0.064 (0.012)
R2 0.126 0.123 -- 0.192 0.193 --
Note: Numbers in parentheses are standard errors. Standard errors are corrected to allow for group effects within survey household clusters. Sample
size in all regressions is 6,326.
aPension income is instrumented with the number of age-eligible women and age-eligible men in the household.
bOther covariates included in regression are 14 province indicators, 3 location indicators (urban, rural, and metro), household size, number of
household members ages 0�5, 6�15, 16�18, 19�21, and 22�24.
Source: All variables are from the World Bank/South African Development Research Unit survey, August�December 1993.
Bertrand, Mullainathan, and Miller 35
ficients on pension receipt in columns 3 and 6 are even more strongly negative
than the ols coefficients in columns 1 and 4. Each extra 100 rand of pension in-
come reduces weekly labor supply of prime-age individuals by about 1.7 hours.12
How large are these effects? For simplicity, assume that the pension is split
equally across all prime-age household members.13 Because there are 4.7 prime-
age people in the average household, the coefficient of �17.07 suggests that a
1000 rand change in individual income reduces hours worked by �17.07 times
4.7 (table 2). Average individual income (computed as household income divided
by number of prime-age people in the household) is 272 rand. Average hours,
conditional on working, equal 41.4.14 Scaling by these gives an elasticity of hours
to income of �17.07 times 4.7 times [(0.272) / (41.4)] equals �.53. Similarly, the
elasticity for employment is computed as �0.099 times 4.7 times [(0.272) / (0.229)]
equals �0.55. These elasticities are large if viewed as pure income effects (see
Imbens and others 1999 for U.S. numbers). The elasticities become even more
strongly negative if the pension is assumed to be split over more household
members.15 One reason for the large magnitude is likely the very low employ-
ment rates in the first place, which make the marginal return to search quite small,
in effect lowering the cost of leisure.
Effects on Men and Women
These regressions are also estimated separately on prime-age African men and
women (table 3). More pension income significantly reduces both working hours
and employment rates among prime-age men. More pension income is also as-
sociated with fewer working hours for women, although the effect is smaller
(�0.01 versus �0.015). Moreover, there is no apparent adjustment of female labor
supply on the extensive margin (panel B, column 4). The only labor supply vari-
able that does not appear to be significantly affected by the presence of eligible
elderly is again female employment status (columns 2 and 5). Although the point
estimate is negative, it is not statistically significant. In the preferred specifica-
tion (columns 3 and 6) the effect on hours worked is much larger for men (2.2)
than for women (1.3). Calculations similar to these (and assuming that men and
women earn similar incomes) yield an elasticity of �0.66 for men and �0.43 for
12. One implication is that household income net of the pension declines when pension income in-
creases. This can be verified in the household-level data by studying the effect of pension income on
total nonpension household income. The iv specification finds that nonpension income goes down by
about 1.05 rand for each extra rand of pension money. Jensen (1998) explores another channel, the
decline in remittance income, through which the social pension can affect nonpension income.
13. Equal sharing among all prime-age individuals does not occur in practice, as is shown later in
the article.
14. The scaling is on hours conditional on working because the effect on the work or not-work decision
will be considered separately.
15. How reasonable is the assumption that prime-age people receive the full pension income? Re-
sults reported later show that women respond less to pension income, suggesting that men get a dispro-
portionate share. Duflo (1999) shows that the social pension improved the anthropometric status of
girls under age five, suggesting that some of the pension income is spent on children.
36
the
Table 3. Effect of Old-Age Pension Income on Working Hours and Employment Status of 16- to world
50-Year-Old African Men and Women
Hours worked Employment status bank
ols ols economic
Pension Pension Pension Pension
uptake eligibility iva uptake eligibility iv
Variableb (1) (2) (3) (4) (5) (6)
review,
Men
Pension income � 1000 �15.13 (1.72) -- �22.48 (2.72) �0.098 (0.034) -- �0.201 (0.056)
Household eligibility -- �8.703 (0.849) -- -- �0.086 (0.018) -- vol.
36 dummy
R2 0.163 0.166 -- 0.234 0.234 -- 17,
Women no.
Pension income � 1000 �10.29 (1.11) -- �13.27 (1.73) �0.018 (0.029) -- �0.023 (0.043)
Household eligibility -- �4.810 (0.646) -- -- �0.013 (0.014) -- 1
dummy
R2 0.107 0.100 -- 0.178 0.178 --
Note: Numbers in parentheses are standard errors. Standard errors are corrected to allow for group effects within survey household
clusters. Sample size is 2,532 for men and 3,794 for women. Other covariates included in the regressions are a quartic in age, a dummy
variable for having completed at least eighth grade, 14 province indicators, 3 location indicators (urban, rural, and metro), household size,
number of household members ages 0�5, 6�15, 16�18, 19�21, and 22�24.
aPension income is instrumented with the number of age-eligible women in the household and the number of age-eligible men in the
household.
Source: All variables are from the World Bank/South African Development Research Unit survey, August�December 1993.
Bertrand, Mullainathan, and Miller 37
women on the hours dimension and elasticities of �0.98 for men and �0.14 for
women on the employment dimension.
In regressions not reported here (but available from the authors) the drop in
employment for men is explored more closely. The regressions examine whether
the "missing" working men have entered a phase of unemployment or whether
they have dropped out of the labor force. No difference in unemployment prob-
abilities is found for eligible and noneligible households. Rather, the missing
working men appear to have left the labor force. Moreover, there is no sign that
the social pension increases the probability of discouragement.
IV. Possible Confounding Effects
The study next looks at the possibility that the estimates of the pension effects
might be biased in that they attribute to the pension the effects of other, unob-
served differences, or that they capture some other behavioral changes induced
by the program rather than a decrease in labor supply.
Direct Effect of the Presence of Elderly in a Household
A primary concern about the results reported here is that individuals living in
pension-eligible households might be systematically different from individuals liv-
ing in pension-ineligible households. For example, the prime-age men and women
living in eligible households are slightly younger than their counterparts in
noneligible households. Furthermore, pension-eligible households are larger on
average. It is conceivable that prime-age men living with older individuals are less
qualified for work, less willing to look for work, or in some other way less likely
to find work. If this is the case, then the estimates of the pension's effect are biased
because they attribute the effect of these unobserved differences to the pension.
Several approaches are used to address this possibility. First, the nonlinearity
in pension receipt as a function of the elder household member's age is exploited
to better separate the pension's effect from these confounding factors. The pen-
sion program rules predict a specific form for these nonlinearities: the presence
of a woman older than 60 or of a man older than 65 should have large effects.
There are no obvious reasons to expect such specific nonlinearities at these two
age thresholds if the estimates are capturing a general impact of the presence of
elderly people on the labor supply of younger household members.
To examine how working hours for prime-age individuals living in a three-
generation household are affected by the presence of elderly in different age
groups, the impact on prime-age labor supply of living with eligible elderly is
first compared with the impact of living with noneligible elderly (a dummy vari-
able is used for the presence in the household of a woman between age 50 and
60 or a man between age 50 and 65). The presence of a noneligible elderly per-
son in the household has neither a statistically nor an economically significant
impact on prime-age working hours (table 4). But as already demonstrated, liv-
ing with an eligible elderly person has a dramatic effect on working hours.
38
the
world
Table 4. Effect of the Presence of Elderly on Hours Worked by 16- to 50-Year-Old Africans
Variable (1) (2)a (3) (4) (5) (6) (7)
bank
Eligible elderly in household �6.79 (0.64) -- -- -- -- -- --
Noneligible elderly �0.46 (0.63) -- -- -- -- -- -- economic
in household
Persons in household 50�55 -- �0.42 (0.50) �0.22 (0.50) �0.40 (0.50) �0.21 (0.50) -- --
(n5055)
Women in household 50�55 -- -- -- -- -- �0.65 (0.67) �0.49 (0.67) review,
(n5055f)
Men in household 50�55 -- -- -- -- -- �0.23 (0.96) �0.03 (0.95)
(n5055m) vol.
38 Persons in household 55�60 -- �0.23 (0.70) �0.09 (0.71) �0.22 (0.70) �0.08 (0.71) -- --
(n5560) 17,
Women in household 55�60 -- -- -- -- -- �0.19 (1.04) �0.04 (1.04)
(n5560f) no.
Men in household 55�60 -- -- -- -- -- �0.62 (0.86) �0.50 (0.87) 1
(n5560m)
Persons in household 60�65 -- �2.54 (0.58) �2.41 (0.58) �2.53 (0.57) �2.41 (0.57) -- --
(n6065)
Women in household 60�65 -- -- -- -- -- �2.95 (0.94) �2.81 (0.93)
(n6065f)
Men in household 60�65 -- -- -- -- -- �1.16 (1.41) �1.03 (1.41)
(n6065m)
n6065m � deviation from �7.47 (3.86) �7.43 (3.88)
eligibility rule in regionb
Persons in household over 65 -- �5.37 (0.51) �5.21 (0.53) -- -- -- --
(n65p)
Persons in household 65�70 -- -- -- �5.17 (0.56) �5.03 (0.72) -- --
(n6570)
Women in household 65�70 -- -- -- -- -- �6.67 (0.86) �6.52 (0.88)
(n6570f)
Men in household 65�70 -- -- -- -- -- �3.85 (1.19) �3.74 (1.20)
(n6570m)
Persons in household over 70 -- -- -- �5.49 (0.56) �5.31 (0.57) -- --
(n70p)
Women in household over 70 -- -- -- -- -- �7.47 (0.72) �7.28 (0.71)
(n70pf)
Men in household over 70 -- -- -- -- -- �2.87 (1.13) �2.76 (1.14)
(n70pm)
Household members over 50 -- -- �1.05 (0.69) -- �1.04 (0.69) -- �0.94 (0.68)
with health problems
R2 0.119 0.124 0.125 0.124 0.125 0.129 0.130
Note: Numbers in parentheses are standard errors. Standard errors are corrected to allow for group effects within survey household clusters. Sample
size in all regressions is 6,326. Other covariates included in regression are a quartic in age, a dummy for sex, a dummy for completion of at least 8th
grade, 14 province indicators, 3 location indicators (urban, rural and metro), household size, number of household members ages 0�5, 6�15, 16�18,
39 19�21, and 22�24.
aTests of equality of coefficient below and above eligibility threshold: Column 2: n5055 = n6065 (p = 0.004), n5055 = n65p (p = 0.000), n5560 =
n6065 (p = 0.009), n5560 = n65p (p = 0.000). Column 3: n5055 = n6065 (p = 0.003), n5055 = n65p (p = 0.000), n5560 = n6065 (p = 0.009); n5560 Bertrand,
= n65p (p = 0.000). Column 4: n5055 = n6065 (p = 0.004), n5055 = n6570 (p = 0.000), n5055 = n70p (p = 0.000), n5560 = n6065 (p = 0.009), n5560
= n6570 (p = 0.000), n5560 = n70p (p = 0.000). Column 5: n5055 = n6065 (p = 0.003), n5055 = n6570 (p = 0.000), n5055 = n70p (p = 0.000), n5560
= n6065 (p = 0.009), n5560 = n6570 (p = 0.000), n5560 = n70p, (p = 0.000). Column 6: n5055f = n6065f (p = 0.020), n5055f = n6570f (p = 0.000),
n5055f = n70pf, (p = 0.000), n5055m = n6570m (p = 0.011), n5055m = n70pm (p = 0.045), n5560f = n6065f, (p = 0.030), n5560f = n6570f (p = Mullainathan,
0.000), n5560f = n70pf (p = 0.000), n5560m = n6570m (0.023), n5560m = n70pm (p = 0.100), n6065m = n6570m (p = 0.131), n6065m = n70pm (p
= 0.379), n5055f = n6065f (p = 0.018), n5055f = n6570f (p = 0.000), n5055f = n70pf. Column 7: (p = 0.000), n5055m = n6570m (p = 0.009), n5055m
= n70pm (p = 0.037), n5560f = n6065f, (p = 0.029), n5560f = n6570f (p = 0.000), n5560f = n70pf (p = 0.000), n5560m = n6570m (p = 0.022), n5560m
= n70pm (p = 0.098), n6065m = n6570m (p = 0.129), n6065m = n70pm (p = 0.378).
bDeviation from eligibility rule in region, the fraction of households with men 60�65 years old and no eligible elderly who are receiving a social
pension in the region. This variable ranges from 0 to 0.67. and
Source: All variables are from the World Bank/South African Development Research Unit survey, August�December 1993.
Miller
39
40 the world bank economic review, vol. 17, no. 1
To further refine this finding, regressors are added for the number of people
in each of the following age categories: 50�55, 55�60, 60�65, and 65 and older
(column 2). The coefficients clearly show a negative effect of the presence of
elderly between 60 and 65 years old and an even stronger negative effect of the
presence of elderly older than 65. On the other hand, the presence of elderly
people between ages 50 and 55 and between ages 55 and 60 seems to have nei-
ther an economically nor a statistically significant impact on working hours
among prime-age individuals. Moreover, the test statistics clearly reject the hy-
pothesis that the pre-eligibility coefficients are equal to the posteligibility coeffi-
cients (see table 4).
Although these results provide some compelling evidence, it is still possible
that the effect of the elderly person's age has an independent, nonlinear effect.
Most notably, the very old are more likely to have health problems and to re-
quire some assistance at home. This may cause prime-age individuals who live
with them to reduce their labor force participation to provide home care.16
The second strategy tries to deal directly with this problem. The survey asks
respondents to list any household member who has been sick or injured over the
past two weeks, "including people who have some form of permanent injury,
disability, or ailment." To investigate whether such health problems display the
same nonlinearity as the pension rule, the probability of having some health
problem was regressed on 10 dummy variables for age and gender groups of
elderly for the entire data set (the results are not detailed here).17 Although people
50�55 years old appear healthier than people older than 55 (this is true for both
men and women), there is no statistically significant difference in the probabil-
ity of being sick for people age 55�60 and people older than 60. Hence, if there
is any age discontinuity in health status, it appears to occur before the age of
pension eligibility.
The number of elderly household members who report health problems are
then included in the employment regression (column 3 of table 4). Though the
coefficient on health problems is negative (each sick elderly person is associated
with an hour less of work) and marginally significant, it does not affect the pen-
sion coefficients. These results exhibit the same discontinuity pattern as they do
in column 2, suggesting that the health status of the elderly does not drive the
results.
Columns 4 and 5 replicate the specifications in columns 2 and 3 but further
break down the number of people older than 65 into number of people age 65�
70 and number of people older than 70. The results are unchanged. The coeffi-
16. At first glance, this story seems inconsistent with the fact that men reduce their work more than
women do. If women provide the main input in home care, they ought to reduce their work hours more.
One could, however, claim that women are expected to both care for the elderly and work, whereas
men will either work or take care of the elderly.
17. The dummy variables are women age 50�55, women age 55�60, women age 60�65, women
older than 70, and the equivalent age groups for men.
Bertrand, Mullainathan, and Miller 41
cients on all the age categories below the eligibility threshold are not statistically
significant different from zero. The coefficients on all the age categories above
the eligibility threshold are significant and negative. Moreover, the test statistics
show that the hypothesis of equality of the coefficients below and above the eli-
gibility threshold can be rejected (see table 4).
The third strategy exploits regional differences in how the pension program
is implemented. In certain areas authorities deviated from the rule that men are
eligible at a later age than women, informally extending the pension to men
between 60 and 65 years old.18 If the results are in fact due to the pension, then
in the regions that deviated from the official rule the presence of men 60�65 would
be expected to affect household labor supply. Because of the informal nature of
the extension, administrative data are not available on which areas altered the
rule, but a proxy can be computed from the data for the fraction of households
with men 60�65 years old and no other age-eligible elderly who report receiving
some pension income.19 This fraction ranges from 0 in the most compliant prov-
ince to 0.67 in the least compliant province.
This fraction is interacted with a dummy variable for the number of men 60�
65 years old in the household (column 6). All the age groups of column 4 are
further disaggregated by gender categories. The results are striking. None of the
pre-eligibility coefficients are statistically different from zero. All the posteligibility
coefficients are significant and negative. The direct effect of number of men 60�
65 years old (the effect in the provinces that do not deviate from the eligibility
rule) is not statistically different from zero. The interaction term between devia-
tion from the eligibility rule and number of men 60�65 years old is negative and
significant. Finally, 10 of the 12 test statistics reject the assumption of equality
between pre-eligibility coefficient and posteligibility coefficient. The same results
hold after controlling for the number of elderly with health problems (column
7). These results suggest that the extension to pre-eligible men does in fact cor-
relate with the pension's estimated effect, bolstering the argument that the re-
sults are not capturing spurious effects of age.
In summary, the results in this section provide multiple pieces of evidence to
suggest that what is being identified is a causal effect of the pension and not an
independent effect associated with living with elderly people.20
18. As Case and Deaton (1998) report, the age differential in pension eligibility is technically un-
constitutional and under revision at the central government level. Certain local authorities might have
already gone ahead with age equalization by 1993.
19. Remember that pension income is observed not at the individual level but at the household level.
20. In a final attempt to account for possible confounding factors associated with the age of the
elderly, the relationship between employment and pension eligibility prior to the bulk of the reform of
the social pension program in South Africa was examined using the 1991 Population Census. This cross-
sectional household survey was conducted prior to the major extension of the social pension to African
households. Although the process of racial equalization of the social pension has been under way since
the early 1990s, only after 1992 were the means tests unified, racial parity in benefits levels achieved,
and new technologies introduced to improve benefit delivery. Thus, although the 1991 census was not
administered prior to the beginning of reform, it was administered at a time when the pension was far
42 the world bank economic review, vol. 17, no. 1
Is the Labor Supply Response Real?
Even taking as given that the regressions are identifying some causal effect of
the pension, there might still be concern that the results are capturing other be-
havioral changes induced by the program, rather than a decrease in labor supply.
Bertrand and others (2000) present a thorough investigation of such alternative
interpretations, which are simply summarized here.
First, they find no evidence that what is being observed is a shift to casual or
farm employment, which might be more difficult to measure than regular forms
of employment. Reported casual working hours actually decline, and the level of
self-employment does not change. The level of home production activities, such
as agricultural crop production or livestock production, does not change either.
There is also no evidence for the related possibility that prime-age individuals living
with pensioners are investing more in human capital. In fact, it is the older rela-
tives of pensioners, not the school-age ones, who show the largest drop in work-
ing hours. Another alternative is that the results are simply picking up on migration
behavior. The pension may make the unemployed more likely to move in with the
pensioners or the employed more likely to move out. There is no indication that
migration patterns and family size are significantly affected by these variables.
V. Distribution of Effects
The results so far provide some evidence of a redistribution of the pension to-
ward prime-age workers in the household. This section pushes the analysis a step
further and asks whether the South African experiment can teach us more about
how resources are allocated and collective labor supply decisions made within
these extended families.
Test of Income Pooling
There are several prominent theoretical models of resource allocation within
households. One, known as the common preference model, assumes that house-
holds are best described as maximizing a single utility function (Samuelson
1956).21 A central result from the common preference model is that money is
money. Which member of the household gets the marginal dollar of nonlabor
income will affect neither the ultimate consumption level nor the leisure choice
less generous and accessible to Africans. No evidence was found that the large negative employment
effects in pension-eligible households (�6.8 percent) using the 1993 survey data are present in the 1991
data. Though there is some negative effect associated with the presence of an age-eligible person in
1991 (not surprising, given that a limited pension program was already in place), the effect is less than
a quarter that of the 1993 pension. These census results are discussed in more detail in Bertrand and
others (2000).
21. The common preference approach can be motivated either through the assumption of a family
consensus, as in Samuelson (1956), or through the assumption of altruistic behavior, as in Becker's
"rotten kid" theorem (Becker 1974, 1981).
Bertrand, Mullainathan, and Miller 43
of each household member. This result holds even in the presence of differential
altruism across individuals. The individuals who get more resources receive the
greatest weight in the joint household utility function.
Another important set of models rejects the idea that families can be reduced
to a single optimizing agent. These models assume that household members have
distinct preferences, and the models look at how bargaining between members
affects the allocation of resources. Most often the bargaining consists of a Pareto
efficient process, such as a Nash bargaining model between the different par-
ties.22 A central feature of these bargaining models is that the strong fungibility
result found in the common preference model no longer holds: who gets the
money matters. The income controlled by each household member influences
the bargained outcome. Moreover, the higher the bargaining power of a house-
hold member, the greater the resources that member will receive.
The social pension program in South Africa provides an unusual opportunity
for an experiment that separates common preference and bargaining models of
the family. As mentioned earlier, the social pension, although in theory means-
tested, is in practice mainly a lump-sum transfer for African households. Hence,
the pension transfer, especially when instrumented for by the presence of age-
eligible elderly, does not depend on earned income or other possible choice vari-
ables in the household resource allocation decision. One can therefore test for
income pooling by asking whether pension transfers made to elderly women have
the same effect on prime-age labor supply as pension transfers made to elderly
men, holding family composition constant.
The findings reported in table 4 already suggest that an elderly woman's pen-
sion income may have a larger negative effect on prime-age labor supply than
an elderly man's pension income. The coefficients on number of women above
the eligibility threshold are systematically larger (in absolute value) than the
coefficients on number of men above the eligibility threshold. When the work-
ing hours of both men and women are regressed on the standard set of geographic,
individual, and family controls, and regressors are added for number of age-
eligible women and age-eligible men, the coefficient on number of eligible women
is more than twice that on number of eligible men (column 1 of table 5).
Such differences are not due to any measurement error in the number of eli-
gible men. Even when accounting for the fact that some men 60�65 years old
also receive pension income in certain South African provinces (column 2), the
coefficient on number of men older than 65 is still only half that on number of
women older than 60.23 Finally, this finding still holds after controlling for the
22. Several researchers, such as Manser and Brown (1980), McElroy and Horney (1981), and
Lundberg and Pollak (1993), have developed cooperative Nash bargaining models of intrahousehold
resource allocation. Chiappori (1992) produces a far more general model that includes all Pareto effi-
cient bargaining models.
23. The coefficient on the interaction term between number of men 60�65 and deviation from the
eligibility rule is insignificant though negative. The rise in standard error is due to the fact that each of
the age categories is not included separately.
44 the world bank economic review, vol. 17, no. 1
Table 5. Old-Age Pension and Pooling of Resources
Prime-age living
with exactly
All prime-age 1 elderly woman
in 3-generation households and 1 elderly man
Variablea (1) (2) (3) (4)
Number of women over 60 �5.02 (0.58) �5.13 (0.58) �5.13 (0.57) �3.89 (1.44)
Number of men over 65 �2.32 (0.87) �2.55 (0.87) �2.54 (0.88) �0.71 (1.47)
Number of men 60�65 -- �1.13 (1.46) �1.12 (1.43) --
Number of men 60�65 � -- �5.31 (4.06) �5.10 (4.05) --
deviation from elibility ruleb
Number of women 16�50 -- -- 0.089 (0.19) --
Number of men 16�50 -- -- �0.14 (0.22) --
Number of women 0�16 -- -- �0.41 (0.17) --
Number of men 0�16 -- -- 0.01 (0.18) --
R2 0.118 0.120 0.122 0.120
Note: Numbers in parentheses are standard errors. Standard errors are corrected to allow for group
effects within survey household clusters. Sample size is 6,326 in columns 1�3 and 1,471 in column 4.
aOther covariates included in all regressions are 14 province indicators and 3 location indicators
(urban, rural, and metro). Also included in columns are a quartic in age, a dummy for sex, a dummy for
having completed at least eighth grade, household size, number of household members ages 0�5, 6�15,
16�18, 19�21, and 22�24.
bFraction of households with men 50�65 years old and no eligible elderly who are receiving a social
pension in the region. This variable ranges from 0 to.67.
Source: All variables are from the World Bank/South African Development Research Unit survey,
August�December 1993.
number of prime-age men, number of prime-age women, number of male chil-
dren and number of female children (column 3). In other words, the fact that
women's pension money reduces labor supply more than men's pension money
cannot be explained by any systematic difference in the number and gender com-
position of the nonelderly in households with eligible women and in households
with eligible men.
This first finding appears inconsistent with pooling of resources within the
household and builds some preliminary support against a common-preference
model of collective labor supply. The marginal dollar of pension income received
by an elderly woman reduces labor supply more than the marginal dollar of
pension income received by an elderly man. However, this first finding is not
conclusive. It does not account for the possibility that the marginal rand of pen-
sion income going to an elderly woman may have to be distributed among a
different set of household members then the marginal dollar of pension money
going to an elderly man.
Of primary concern here is the possibility that old women might have a lower
weighting than old men in a household utility function. If that is the case, and
assuming that households with eligible women have more elderly women than
households with eligible men (a very likely event), the finding could still be recon-
ciled with the common preference model. To deal with this concern the sample is
Bertrand, Mullainathan, and Miller 45
restricted to the set of households that have exactly one elderly woman (older than
50) and one elderly man (also older than 50).24 In this subset of households the
marginal rand of pension income, whether from a female or a male pensioner, will
be reallocated among a fixed number of elderly of each gender. Replicating the
specification of column 1 on that subset of households still yields a much stronger
negative labor supply response for pension money going to women than pension
money going to men (column 4 of table 5). The marginal rand of pension income
going to a female pensioner reduces labor supply by about three times as much as
the marginal rand of pension income going to a male pensioner. The coefficient
on number of age-eligible men is, however, less precisely estimated.
Although the number and gender composition of the elderly are forced to be
the same in this restricted sample, it could be that the number and gender com-
position of the nonelderly systematically varies with the gender of the pensioner.
For the subset of households with exactly one woman older than 50 or one man
older than 50, no statistically significant difference was found in the number and
gender of prime-age individuals between households with a female pensioner and
households with a male pensioner. However, households with a female pensioner
had slightly more children than households with a male pensioner. This last dif-
ference suggests that the marginal rand of pension money needs to be split among
slightly more individuals when the pensioner is a woman. Hence, under the com-
mon preference model, this could only lead the coefficient on eligible women to
be smaller in absolute value than the coefficient on eligible men if children re-
ceive any weighting in the family utility function. This is exactly the opposite of
the results found.
In summary, the results in table 5 strongly point toward rejecting the income-
pooling hypothesis for African extended families. A marginal rand of pension
income has a drastically different effect on prime-age labor supply depending
on whether the pension-earning person is a man or a woman. This still holds
true when households differ only in the gender of their pensioners and not in the
number, age, and gender composition of their members.
Who Benefits from the Old-Age Pension?
Testing for income pooling is only one way to look at how money is distributed
inside the household. Another would be to examine who are the biggest benefi-
ciaries of the reallocation of resources. Is the pension money evenly distributed
among all the prime-age individuals in the extended family or are certain family
members able to reduce their labor supply more than others? Answering this
question involves estimating the standard regression of hours worked on the
pension variable, but this time interacting the pension variable with several de-
mographic characteristic (table 6).25
24. There are few households that contain two or more elderly persons of both genders.
25. For reasons of space results are reported in table 6, mainly for the iv specification.
46
the
world
Table 6. Distribution of the Effect of Old-Age Pension Income on Working Hours of 16- to 50-Year-Old Africans
Variablea (1) (2)b (3) (4) (5) (6) (7)
bank
Pension income � 1000 �21.04 (2.51) -- �14.09 (1.85) �15.55 (1.63) 25.48 (3.45) 39.35 (12.37) 20.79 (3.66)
Pension income � 1000 � female 9.05 (2.21) -- -- -- -- -- 7.10 (2.47) economic
Women over 60 -- �6.98 (0.81) -- -- -- -- --
Men over 65 -- �2.73 (1.03) -- -- -- -- --
Women over 60 � female -- 3.23 (0.75) -- -- -- -- --
Men over 65 � female -- 0.71 (0.89) -- -- -- -- -- review,
Pension income � 1000 � -- -- �7.42 (3.02) -- -- -- --
4th grade or less
Pension income � 1000 � matric -- -- -- �0.07 (3.93) -- -- -- vol.
46 or more
Pension income � 1000 � age -- -- -- -- �1.53 (0.14) �2.54 (0.91) �1.49 (0.14) 17,
Pension income � 1000 � age2 -- -- -- -- -- 0.02 (0.02) --
no.
Pension income � 1000 � oldest -- -- -- -- -- -- �9.08 (5.00)
prime-age man in household 1
Note: Numbers in parentheses are standard errors. Standard errors are corrected to allow for group effects within survey household clusters. Sample size
is 6,326 in columns 1�6 and 6,189 in column 7.
aOther covariates included in regression are a quartic in age, a dummy variable for gender, a dummy variable for having completed at least 8th grade, 14
province indicators, 3 location indicators (urban, rural, and metro), household size, number of household membersages 0�5, 6�15, 16�18, 19�21, and 22�
24. Columns 3, 4, and 7 also include a dummy variable for "4th grade or less," a dummy variable for "matric or more," and a dummy variable for "oldest
prime-age man in the household."
b All columns except column 2 represent iv results. In the iv specifications, pension income and the interactions of pension income with the other variables
of interest are instrumented.
Source: All variables are from the World Bank/South African Development Research Unit survey, August�December 1993.
Bertrand, Mullainathan, and Miller 47
As the results in table 3 have already suggested, the pension reduces the labor
supply of prime-age men more than that of prime-age women. The effect of the
social pension on women's labor supply is about half the effect on men's (col-
umn 1 of table 6), implying that prime-age women benefit less from the social
pension than do prime-age men. To see whether this effect depends on the gen-
der of the pensioner, the female dummy variable is interacted with the number
of eligible women and the number of eligible men in the household (column 2).
The presence of an additional male pensioner in a household does not have a
statistically different effect on male and female labor supply, whereas the pres-
ence of an additional female pensioner benefits prime-age men more than it does
prime-age women. That means that the greater effect of the pension on prime-
age men occurs only when the pensioner is a woman.
Education might also influence who benefits more from the pension transfer.
On one hand, individuals with higher educational attainment presumably have
more outside options, which could increase their threat points when bargaining
over resources with other family members. On the other hand, at a given level of
redistribution, individuals with the lowest market wages may give up their job
first. If educational attainment is positively correlated with market wage, the least
educated workers would be expected to reduce their labor supply most. The
differential effect of the pension by education group thus appears to be an em-
pirical question. The prime-age men and women in the sample who have not
completed fourth grade (about a quarter in each group) reduce their labor sup-
ply by about 50 percent more than the individuals who have completed at least
the fourth grade (column 3). There is no difference in labor supply response,
however, between individuals who have completed at least the matric (10th grade)
and those who have not. Thus, it appears that the labor supply response is stron-
gest among the least skilled, probably because they face such unattractive labor
market options to start with.
Age could also affect the size of the labor supply response of prime-age house-
hold members. The social pension depresses labor supply more as prime-age men
get older (column 5). Allowing for a quadratic relation between pension money
and age enables an assessment of whether the effect peaks at any point over the
range of working ages. The effect of age appears mostly linear and does not peak
before 50 years of age (column 6).26
Another question is whether it is absolute or relative age that affects intra-
household redistribution. More precisely, does the special position that oldest
sons are believed to hold inside the family result in the oldest prime-age man
in the household receiving more pension money than other household mem-
bers? After the differential effects of the pension by age and gender are con-
trolled for, the results show that the oldest man in the household reduces his
26. These results also undercut the argument that household members reduce their labor supply to
get an education.
48 the world bank economic review, vol. 17, no. 1
labor supply more than other household members (column 7).27 These results
support the view that oldest sons receive more resources in extended families.
After the direct effect of own age and gender on resource distribution are ac-
counted for, the results show that the oldest man reduces his labor supply by
about 50 percent more than other men in the household and about 70 percent
more than women.
The results in this section on the distribution of effects can be understood in
light of the bargaining models of household resource allocations. First, the ob-
served differences in redistribution can be attributed to differences in bargain-
ing power. Men's labor supply responds more to pension income because men
have more power inside the household. That the male�female differential is largest
when the pensioner is a women is suggestive of a situation in which dominant
males capture resources. When the pensioner is a man, the ability of prime-age
men to capture resources is diminished and so is the male�female differential in
labor supply response. The age results are in line with this picture. The oldest
male seems most capable of capturing household resources.
Alternatively, differences in altruism could explain the patterns observed in
table 6. Perhaps pensioners care more about men. To fit the results here, female
pensioners would have to care the most about prime-age male household mem-
bers. Moreover, pensioners' altruism would have to be strongest toward their
oldest prime-age male children. Even if this pattern of altruism does not seem
particularly intuitive, it nevertheless provides another lens for interpreting the
findings.
VI. Conclusion
With improving health conditions and lengthened life expectancy in many de-
veloping economies, governments will soon have to introduce full-fledged social
programs to provide for the needs of a growing elderly population. Before simply
replicating the type of programs that industrial countries have installed, policy-
makers in developing economies would do well to consider how different living
arrangements could interfere with their social objectives. Though the elderly in
industrial countries may often live on their own, multigeneration households
prevail in developing countries. The South African pension program provides a
way to understand the effects of such targeted programs when extended family
links are strong.
The South African government's pension program was introduced as a way
to improve living conditions for older individuals who are no longer in the labor
force and who do not have access to a private pension. The vast majority of the
older Africans in South Africa participate in this program. This article provides
some evidence that, in practice, at least part of the cash transfers targeted for the
27. The sample size is slightly smaller here than in the basic regression because the 2 percent of
households that have only one prime-age individual are excluded from the sample.
Bertrand, Mullainathan, and Miller 49
elderly ends up in the hands of a group that was not originally targeted: prime-
age men and women who live with the pensioners. The results reported here
indicate that African household members 16�50 years old reduce their labor
supply when they live with pension beneficiaries. Hence, because of intra-
household redistribution, a program designed for a group that is out of the labor
force unexpectedly altered the labor supply of a nontargeted group.
Moreover, the study relates this labor supply response to standard theories of
intrahousehold resource allocation and collective labor supply choice. The dif-
ferent labor supply impact of money from male and female pensioners suggests
that a common preference model of family labor supply cannot adequately de-
scribe the results and that some amount of bargaining takes place within these
households. In general, older prime-age men, in particular the oldest prime-age
man in a household, appear to be the biggest beneficiaries of the pension. Within
the set of bargaining models of intrahousehold allocation of resources, this find-
ing could be interpreted as evidence that these men have relatively more bar-
gaining power or are being cared for more by other family members.
References
Altonji, Joseph, Fumio Hayashi, and Laurence Kotlikoff. 1992. "Is the Extended Family
Altruistically Linked? Direct Tests Using Micro Data." American Economic Review
82(5):1177�98.
Ardington, Libby, and Frances Lund. 1994. "Pensions and Development: The Social
Security System as a Complementary Track to Programs of Reconstruction and De-
velopment." University of Natal Centre for Social and Development Studies, Durban.
Becker, Gary. 1974. "A Theory of Social Interaction." Journal of Political Economy
82(6):1063�94.
------. 1981. A Treatise on the Family. Cambridge, Mass.: Harvard University Press.
Bertrand, Marianne, Sendhil Mullainathan, and Douglas Miller. 2000. "Public Policy
and Extended Families: Evidence from South Africa." Working Paper, University of
Chicago Graduate School of Business, Chicago.
Browning, Martin, Francois Bourguignon, Pierre-Andr� Chicappori, and Val�rie Lechene.
1994. "Income and Outcomes: A Structural Model of Intrahousehold Allocation."
Journal of Political Economy 102(6):1067�96.
Caelers, Di. 1998. "Noah Saves Khayelitsha's Old Folk." Independent, June 16.
Case, Anne, and Angus Deaton. 1998. "Large Cash Transfers to the Elderly in South
Africa." Economic Journal 108(450):1330�61.
Chiappori, Pierre-Andre. 1992. "Collective Labor Supply and Welfare." Journal of Po-
litical Economy 100(3):437�67.
Duflo, Esther. 1999. "Child Health and Household Resources in South Africa: Evidence
from the Old Age Pension Program." Working Paper, Massachussetts Institute of
Technology, Cambridge, Mass.
Imbens, Guido, Donald Rubin, and Bruce Sacerdote. 1999. "Estimating the Effects of
Unearned Income on Labor Supply, Earnings, Savings, and Consumption: Evidence
from a Survey of Lottery Players." nber Working Paper 7001. National Bureau of
Economic Research, Cambridge, Mass.
50 the world bank economic review, vol. 17, no. 1
Jensen, Robert. 1998. "Public Transfers, Private Transfers and the `Crowding Out'
Hypothesis: Evidence from South Africa." Working Paper, Harvard University, John
F. Kennedy School of Government, Cambridge, Mass.
Lund, Frances. 1992. "State Social Benefits in South Africa." International Social Security
Review 46(1):5�25.
Lundberg, Shelley, and Robert Pollak. 1993. "Separate Spheres Bargaining and the
Marriage Market." Journal of Political Economy 101(6):988�1010.
Lundberg, Shelley, and Robert Pollak. 1996. "Bargaining and Distribution in Marriage."
Journal of Economic Perspectives 10(4):139�58.
Manser, Marilyn, and Murray Brown. 1980. "Marriage and Household Decision Making:
A Bargaining Analysis." International Economic Review 21(1):31�44.
McElroy, Marjorie, and Mary Jean Horney. 1981. "Nash-Bargained Household Deci-
sions: Towards a Generalization of the Theory of Demand." International Economic
Review 22(2):333�49.
Ngoro, Blackman. 1998. "Retiring to Life in Shackland." Independent, July 20.
Samuelson, Paul. 1956. "Social Indifference Curves." Quarterly Journal of Economics
70(1):1�22.
Subramanian, Shankar, and Angus Deaton. 1991. "Gender Effects in Indian Consump-
tion Patterns." Sarvekshana 14(4):1�12.
Van der Berg, Servaes. 1994. "Issues in South African Social Security." World Bank,
Washington, D.C.
the world bank economic review, vol. 17, no. 1 51�88
Economic, Demographic, and Institutional
Determinants of Life Insurance
Consumption across Countries
Thorsten Beck and Ian Webb
Life insurance has become an increasingly important part of the financial sector over
the past 40 years, providing a range of financial services for consumers and becoming
a major source of investment in the capital market. But what drives the large variation
in life insurance consumption across countries remains unclear. Using a panel with data
aggregated at different frequencies for 68 economies in 1961�2000, this article finds
that economic indicators--such as inflation, income per capita, and banking sector
development--and religious and institutional indicators are the most robust predic-
tors of the use of life insurance. Education, life expectancy, the young dependency ratio,
and the size of the social security system appear to have no robust association with life
insurance consumption. The results highlight the importance of price stability and
banking sector development in fully realizing the savings and investment functions of
life insurance in an economy.
Life insurance companies play an increasingly important role in the financial
sector. In 1980�85 the total assets of life insurance companies accounted for only
11 percent of gross domestic product (gdp) for a sample of 13 countries for which
data are available, but in 1995�97 they accounted for 28 percent of gdp in the
same countries. This greater importance is also reflected in the business volume
of life insurers. While life insurance penetration--the ratio of premium volume
to gdp--was 1.2 percent in 1961�65 for a sample of 19 countries for which data
are available, it reached 4.2 percent in 1996�2000 in these countries.
This increased importance of life insurance as a provider of financial services
and of investment funds in capital markets is especially pronounced for devel-
oped economies, whereas life insurance consumption remains low in many de-
veloping economies. But even among developing countries there are striking
differences. The penetration ratio in South Africa was 12.7 percent in 1996�2000,
but it was less than 0.01 percent in the Syrian Arab Republic. The large varia-
Thorsten Beck is with the Development Research Group at the World Bank; his e-mail address is
tbeck@worldbank.org. Ian Webb is with the International Insurance Foundation; his e-mail address is
webb@iifdc.org. The authors are grateful to Robert Cull, Lisa Gardner, Harold Skipper Jr., partici-
pants in the Finance Forum at the World Bank in June 2002, three anonymous referees, and the editor
for useful comments and discussions. Any remaining errors are the authors'.
DOI: 10.1093/wber/lhg011
� 2003 The International Bank for Reconstruction and Development / THE WORLD BANK
51
52 the world bank economic review, vol. 17, no. 1
tion in the use of life insurance across countries raises questions about what causes
this variation and thus what determines life insurance consumption.
Life insurance provides individuals and the economy with several important
financial services. First, life insurance products encourage long-term saving and
the reinvestment of substantial sums in public and private sector projects. By
leveraging their role as financial intermediaries, life insurers have become a key
source of long-term finance, encouraging the development of capital markets
(Catalan and others 2000; Impavido and Musalem 2000).1 Indeed, several stud-
ies have found evidence that the development of the insurance sector is related
to economic growth (Ward and Zurbruegg 2000; Webb 2000; Soo 1996). Sec-
ond, in the face of growing urbanization, population mobility, and formaliza-
tion of economic relationships between individuals, families, and communities,
life insurance has taken on increasing importance as a way for individuals and
families to manage income risk.
The importance of life insurance for economic and financial development
directs us to investigate which economic, demographic, and institutional factors
give rise to a vibrant life insurance market. Several studies have identified a core
set of socioeconomic determinants as good predictors of life insurance consump-
tion. But the relatively limited data samples and the different measures of con-
sumption used in these studies have limited their scope and made it difficult to
generalize from their conclusions.
In this article we improve on the existing literature in several ways. First, we
use a new data set that significantly extends the coverage of economies and pe-
riods. Previous cross-sectional and panel studies have been limited in depth or
in breadth.2 The new data set spans 68 economies over the period 1961�2000
and includes aggregate data at different frequencies.
Second, panel analysis allows us to exploit both cross-country and time-series
variation in life insurance consumption and its potential determinants. We can
thus better assess what has driven the rapid increase in life insurance consump-
tion over the past four decades. At the same time cross-sectional analysis allows
us to analyze the effect of time-invariant determinants and control for biases
induced by reverse causation and simultaneity.
Third, by using several alternative measures of life insurance consumption, we
provide additional depth and robustness to the results. Life insurance premiums
and life insurance in force--the outstanding face amounts plus dividend additions
of life insurance policies--measure different aspects of life insurance consumption.
Finally, we introduce a new measure for exploring the role of life insurance in
the economy--its relative weight in individual savings portfolios. This indicator
1. For more on the economic and social importance of life insurance, especially in developing coun-
tries, see unctad (1982), one of the first studies in this area.
2. Browne and Kim (1993) use data for 45 countries for 1987, and Outreville (1996) data for 48
countries for 1986. Truett and Truett (1990) produce estimates for two countries, Mexico and the United
States, for 1960�82, and Beenstock and others (1986) estimates for 10 oecd countries for 1970�81.
Beck and Webb 53
measures the weight of life insurance premiums in the private savings in an
economy.
The results are expected to help policymakers understand what drives the
supply of and demand for life insurance. They may also help design strategies
for developing nascent life insurance markets and extending their benefits to more
countries.
I. Measuring Life Insurance Consumption across Countries
Life insurance policies are financial products that offer two main services: in-
come replacement for premature death and a long-term savings instrument. There
are a multitude of types of policies, each offering the consumer different cover-
age options and investment choices, but they can be broken down into two gen-
eral categories: those offering mortality coverage only and those combining
mortality coverage with a savings component. Policies in the first category are
generally referred to in the United States and many other countries as term poli-
cies. Those in the second category are known as whole life, universal life, vari-
able life, endowment, and by a variety of other names. Policies in the second
category typically earn interest, which is returned to the consumer through policy
dividends, cash values on termination of the policy, or endowment sums on
maturation of the policy. These policies incorporate varying amounts of mor-
tality coverage while generally offering a substantial savings component.
In addition to these two categories, life insurers also sell annuity policies.
Annuities are contractual arrangements whereby in return for a lump sum or
periodic payments until annuitization, the insurer promises to make periodic
payments to the insured, often until his or her death. Insurers providing annu-
ities thus undertake risks associated with longevity of the insured.
Because the different measures of life insurance consumption used in our
empirical analysis aggregate both categories of life insurance policies as well as
annuity policies, we cannot distinguish between the demand for and supply of
mortality risk coverage, longevity risk coverage, and savings through life insur-
ance. This aggregation in the data produces a bias against finding significant
relationships (see Browne and Kim 1993, note 1). Significant relationships be-
tween the variables hypothesized to affect insurance consumption and the amount
consumed are therefore likely to signal added robustness in the results.
Life insurance penetration, defined as the ratio of premium volume to gdp,
measures insurance activity relative to the size of the economy. Because it is the
product of quantity and price, it is not a perfect measure of consumption. A larger
premium volume might reflect a larger quantity, a higher price, or a difference
in the mix of mortality risk, savings, and annuity elements purchased. Lack of
competition and costly regulation might increase the amount spent on insurance
by raising its price, without implying higher insurance consumption.
Life insurance density, our second indicator of life insurance consumption, is
defined as premiums per capita. This measure shows how much each inhabitant
54 the world bank economic review, vol. 17, no. 1
of a country spends on insurance on average, expressed in constant dollars.3
Although both life insurance penetration and life insurance density use gross
premiums, important differences remain between the two measures: life insur-
ance penetration measures life insurance consumption relative to the size of the
economy, whereas life insurance density compares life insurance consumption
across countries without adjusting for income. Consumers who purchase life
insurance policies to insure their dependents against mortality risk will poten-
tially buy more coverage and thus a higher face value in richer countries, be-
cause the death benefit has to replace a larger income. We therefore expect life
insurance density to be more income elastic than life insurance penetration.
Because life insurance policies are just as much a savings product as they are
an insurance product, we can relate the total premiums to private savings rather
than income. This implies a portfolio rather than an income approach, treating
life insurance policies as one of several assets from which investors can choose.
We therefore construct the measure life insurance in private savings, equal to
total premiums divided by private savings, to indicate the share of private sav-
ings that the inhabitants of a country invest in life insurance policies.4 Because
of data limitations, this indicator is available only for 1970�95.
Our last measure of life insurance consumption is life insurance in force to
GDP, equal to the sum of the face amounts plus dividend additions of life insur-
ance policies outstanding as a share of gdp. It is a measure of mortality risk
underwritten plus savings accumulated. Life insurance in force thus includes both
the cash value of policies, associated with the savings component of life insur-
ance policies, and the net amount of risk faced by life insurers. Unlike the other
three indicators, life insurance in force to gdp does not include price and so
measures only quantity. As a result of data limitations, this indicator is avail-
able only for 1961�94.
The mortality risk, savings, and annuity components have different weights
in the premium and stock measures. For a given structure of the insurance mar-
3. We also calculate an alternative measure of life insurance density using international real dollars.
Specifically, rather than applying exchange rates, the local currency premiums are multiplied by the
purchasing power parity (ppp) conversion factor, defined as the number of units of a country's cur-
rency required to buy the same amount of goods and services in the domestic market as one U.S. dollar
would buy in the United States. Using ppp conversion factors is preferable to using exchange rates,
because exchange rates are distorted by differences in exchange rate regimes. Moreover, ppp conversion
factors take into account the fact that the price of nontraded goods relative to traded goods increases with
the income level of an economy. Because the death benefit of life insurance policies has to cover the typi-
cal household spending on both traded and nontraded goods, using exchange rates biases the insurance
density of developing economies downward. But because data on the ppp conversion factor are available
only for 1975�2000, the insurance densities in international real dollars are constrained to this period.
All the regressions were run using this alternative indicator of life insurance density without significant
differences, so we report only results with the general measure available over a longer period.
4. According to the United Nations System of National Accounts, life insurance premiums that imply
claims of policyholders on insurance companies' technical reserves are treated as savings, whereas in-
surers' costs and profits are part of consumption. See United Nations Statistics Division (1993).
Beck and Webb 55
ket, the mortality risk component, as measured by the net amount of risk, has a
stronger weight in life insurance in force to gdp than in the other three mea-
sures. In most (but not all) countries life insurance in force does not include
annuities (see Browne and Kim 1993).
Life insurance consumption varies widely across economies. While Syrians
spent less than US$1 a year on life insurance services in 1996�2000, Japanese
spent more than US$3200. Ecuadorians invested less than 1 percent of their total
savings in life insurance policies in 1991�95, and British citizens invested more
than 40 percent in 1986�90. Similarly, life insurance in force was less than 0.1
percent of gdp for Greece in 1976�80, but it reached nearly 400 percent of gdp
for Japan in 1991�95. There are large correlations between all three measures
of life insurance consumption that are significant at the 1 percent level (tables 1
and 2).
II. Determinants of Life Insurance Consumption
In this section we describe the theoretical underpinnings of our empirical tests
and different factors hypothesized to drive the demand for and supply of life
insurance policies.5
Theoretical Underpinnings
Yaari (1965) and Hakansson (1969) were the first to develop a theoretical frame-
work to explain the demand for life insurance. In this framework the demand
for life insurance is attributed to a person's desire to bequeath funds to depen-
dents and provide income for retirement. The consumer maximizes lifetime util-
ity subject to a vector of interest rates and a vector of prices, including insurance
premium rates. This framework posits that the demand for life insurance is a
function of wealth, expected income over a person's lifetime, interest rates, the
cost of life insurance policies (administrative costs), and the assumed subjective
discount rate for current over future consumption.
Lewis (1989) extends this framework by explicitly incorporating the prefer-
ences of the dependents and beneficiaries into the model. Specifically, he derives
the demand for life insurance as a maximization problem of the beneficiaries,
the spouse, and the offspring of the policyholder. Deriving utility maximization
by the spouse and offspring separately and assuming no bequest by the policy-
holder and an isoelastic utility function, Lewis shows that total life insurance
demand can be written as
(1) (1 � lp)F = max{[(1 � lp) / l(1 � p)]1/dTC � W,0}
where l is the policy loading factor (the ratio of the cost of the insurance to its
actuarial value), p the probability of the primary wage earner's death, F the face
5. For an excellent overview of the potential determinants of the demand for and supply of life in-
surance products, see Skipper and Black (2000, chap. 3).
56 the world bank economic review, vol. 17, no. 1
Table 1. Descriptive Statistics
Variable Mean Median SD Maximum Minimum Observations
Life insurance penetration 1.69 1.03 1.97 12.69 0.00 322
Life insurance density 264.51 68.88 442.45 3275.39 0.14 322
Life insurance in 7.64 4.64 8.24 44.90 0.00 203
private savings
Life insurance in force 56.25 29.85 60.69 398.43 0.09 216
to gdp
gdp per capita 9463 4393 10,090 45,061 193 451
Young dependency ratio 55.14 50.64 23.02 107.26 21.41 451
Old dependency ratio 12.52 9.64 6.44 27.65 4.50 451
Life expectancy 68.17 70.71 8.07 80.48 41.63 451
Schooling 5.76 5.60 2.72 12.18 0.63 451
Inflation rate 14.37 7.32 25.63 222.33 �0.10 451
Banking sector development 47.29 38.62 32.65 180.88 5.41 451
Gini index 37.41 34.89 9.61 61.88 20.46 221
Urbanization 60.26 61.00 21.63 100.00 8.11 451
Social security 12.13 9.57 8.98 38.26 0.46 343
Real interest rate 26.44 1.80 260.74 3686.98 �46.13 402
Expected inflation rate 14.31 7.41 25.52 232.85 �0.03 451
Permanent income 9450 4329 10,172 51,429 176 451
Secondary enrollment 67.71 69.51 29.58 152.84 7.67 399
Private savings rate 20.54 20.95 5.93 37.45 2.81 264
Revolutions and coups 0.17 0.00 0.34 2.60 0.00 312
Human development index 0.75 0.77 0.13 0.93 0.35 304
Rule of law 4.13 4.00 1.53 6.00 1.00 245
Inflation volatility 6.94 2.79 16.50 169.73 0.21 451
Institutional development 0.48 0.54 0.78 �1.33 1.72 69
Catholic 41.04 29.80 40.03 0 96.9 69
Muslim 13.12 0.55 29.28 0 99.4 69
Protestant 14.64 2.60 25.26 0 97.8 69
British legal origin 0.26 0.00 0.44 0 1 69
French legal origin 0.45 0.00 0.50 0 1 69
Socialist legal origin 0.12 0.00 0.32 0 1 69
German legal origin 0.09 0.00 0.28 0 1 69
Scandinavian legal origin 0.07 0.00 0.26 0 1 69
Good crops 1.15 1.06 0.32 0.65 2.44 65
Source: Appendix table A-1.
value of all life insurance written on the primary wage earner's life, d a measure
of the beneficiaries' relative risk aversion, TC the present value of consumption
of each offspring until he or she leaves the household and of the spouse over his
or her predicted remaining life span, and W the household's net wealth. Demand
for life insurance increases with the probability of the primary wage earner's
death, the present value of the beneficiaries' consumption, and the degree of risk
aversion. It decreases with the policy loading factor and the household's wealth.
But life insurance consumption is not driven only by consumer demand. Im-
portant supply-side factors affect the availability and price of life insurance. In-
surance companies need human and information resources to effectively measure
Table 2. Correlations
Life Life insurance Life Young Old
Life insurance insurance in private insurance in gdp per dependency dependency Life Inflation
Variable penetration density savings force to gdp capita ratio ratio expectancy Schooling rate
Life insurance density 0.7881*** 1.0000
Life insurance in 0.9357*** 0.6918*** 1.0000
private savings
Life insurance in force 0.7729*** 0.7434*** 0.6444*** 1.0000
to gdp
gdp per capita 0.5219*** 0.7481*** 0.4241*** 0.4870*** 1.0000
Young dependency ratio �0.3673*** �0.4667*** �0.3511*** �0.3949*** �0.7297*** 1.0000
Old dependency ratio 0.2885*** 0.4680*** 0.3261*** 0.2348*** 0.7763*** �0.8278*** 1.0000
Life expectancy 0.2784*** 0.4673*** 0.2834*** 0.4169*** 0.6912*** �0.8310*** 0.7159*** 1.0000
Schooling 0.5001*** 0.5471*** 0.5181*** 0.5724*** 0.7330*** �0.7980*** 0.7217*** 0.7882*** 1.0000
Inflation rate �0.2594*** �0.2274*** �0.2553*** �0.1769*** �0.2022*** 0.0339 �0.0973** �0.0424 �0.0523 1.0000
Banking sector 0.5031*** 0.5866*** 0.3916*** 0.4462*** 0.6748*** �0.6150*** 0.5086*** 0.5622*** 0.5245*** �0.2148***
development
**Significant at the 5 percent level.
***Significant at the 1 percent level.
Source: Authors' calculations.
58 the world bank economic review, vol. 17, no. 1
the pricing and reserve requirements for products as well as adequate investment
opportunities in financial markets. Adequate protection of property rights and
effective enforcement of contracts also facilitate the investment function of life
insurers. These supply factors are expected to affect the costs of life insurance
products and might therefore be represented by the policy loading factor in the
Lewis model.
Attempts have been made to model the relationship between the supply of
and demand for life insurance separately, but data limitations have restricted
empirical testing of the models (see Beenstock and others 1986). The available
data do not allow us to distinguish between supply and demand. Moreover,
premium data do not allow us to observe the actual amount of insurance cover-
age purchased, as they are a combined measure of price and coverage. Unless
the price is constant across countries, which is unlikely, assuming that the pre-
mium is equivalent to the amount of coverage would introduce a source of noise
in our estimations. But using the variable often employed to proxy price (premi-
ums over life insurance in force) requires a troublesome assumption--that the
mix of policies remains constant across countries and over time.6
Price is undoubtedly an important determinant of the consumption of life
insurance, however, and leaving it out may subject the empirical testing to omit-
ted variable bias. We address this problem in two ways. First, we assume that
the price is a function of several supply-side factors. Varying levels of urbaniza-
tion, monetary stability, institutional development, political stability, and banking
sector development all affect insurers' ability to provide cost-effective insurance.
Second, we use panel estimation techniques that eliminate biases due to omitted
variables, such as the price variable in our model.
In the following sections we describe variables that may be linked to the de-
mand function described by Lewis (1989) as well as several supply factors that
might proxy for the policy loading factor. While the Lewis model focuses on the
mortality risk component of life insurance policies, we link the different deter-
minants to the savings and annuity components of life insurance policies as well.
The portfolio approach underlying life insurance in private savings adds another
dimension to the discussion.
Demographic Determinants
A higher young dependency ratio (the ratio of young dependents to the working-
age population) is assumed to increase the demand for mortality coverage and
decrease the demand for savings through life insurance and annuities (table 3;
see table A-1 for the construction and sources of the variables). A larger share of
dependents in the population means a higher total present value of consump-
tion of the beneficiaries of those insured--and therefore a higher demand for
life insurance that provides dependents with payments in the event of the pre-
6. Browne and Kim (1993) use such a price variable, but they note the bias introduced by different
compositions of the overall insurance portfolio across countries.
Beck and Webb 59
Table 3. Determinants of Life Insurance Consumption across Countries:
Expected Results of the Regression Analysis
Expected
Expected effect Expected effect Expected effect effect on all
on savings on mortality on annuity components
Variable component risk component component combined
Demographic variables
Young dependency ratio � + � Ambiguous
Old dependency ratio + � + Ambiguous
Life expectancy + � + Ambiguous
Schooling + + + +
Religion (Muslim) � � � �
Urbanization + + + +
Economic variables
Income + + + +
Private savings rate Ambiguous No effect Ambiguous Ambiguous
Inflation rate � � � �
Inflation volatility � � � �
Real interest rate + + + +
Banking sector development + + + +
Social security � � � �
Gini index Ambiguous Ambiguous Ambiguous Ambiguous
Institutional variables
Rule of law + + + +
Revolutions and coups � � � �
Institutional development + + + +
Note: This table assumes the division of life insurance consumption into the savings, mortality risk,
and annuity components.
Source: See section on determinants of life insurance consumption.
mature death of the primary wage earner (this would result in a higher TC in
equation 1). A high young dependency ratio also means that a large share of the
population is too young to consider saving for retirement--and thus implies lower
demand for savings through life insurance products. Beenstock and others (1986),
Browne and Kim (1993), and Truett and Truett (1990) find that the young de-
pendency ratio is positively correlated with life insurance penetration. Given the
opposite effects of the young dependency ratio on the mortality and savings
components of life insurance, however, we predict that a higher young depen-
dency ratio is ambiguously correlated with life insurance.
A higher old dependency ratio (the ratio of old dependents to the working-
age population) is assumed to increase the demand for the savings and annuity
components and decrease the demand for the mortality risk component of life
insurance. We conjecture that in countries in which a larger share of the popu-
lation is retired, savings through life insurance policies and protection against
outliving one's retirement income gain importance, whereas insurance against
the risk of the primary wage earner's death loses importance. The overall effect
of the old dependency ratio is therefore predicted to be ambiguous.
60 the world bank economic review, vol. 17, no. 1
Societies with a longer life expectancy should have lower mortality coverage
costs, lower perceived need for mortality coverage, but higher savings through
life insurance vehicles and more demand for annuities (a longer life expectancy
would be reflected by a lower p in equation 1). This would imply an ambiguous
correlation with the demand for life insurance products (compare Beenstock and
others 1986). Earlier studies have found life expectancy to be positively corre-
lated with life insurance penetration (Beenstock and others 1986; Outreville
1996).
We expect that a higher level of education in a population will be positively
correlated with the demand for any type of life insurance product. A higher level
of education may increase people's ability to understand the benefits of risk
management and long-term savings--and therefore increase their risk aversion
(this would be reflected by a lower d in equation 1).7 Education may also in-
crease the demand for pure death protection by lengthening the period of de-
pendency as well as by increasing the human capital of--and so the value to be
protected in--the primary wage earner (this would be reflected by a higher TC
in equation 1). But a positive relationship between education and life insurance
might also indicate that better access to long-term savings and insurance instru-
ments encourages access to higher education.8 Truett and Truett (1990) and
Browne and Kim (1993) find a positive relationship between life insurance con-
sumption and the level of education. To measure the education level, we use the
average years of schooling in the population over age 25 and the gross second-
ary enrollment ratio.
The religious inclination of a population may affect its risk aversion and its
attitude toward the institutional arrangements of insurance (this would be re-
flected by cross-country variation in d in equation 1). Religious opposition to
life insurance, though stronger in European countries before the 19th century,
persists in several Islamic countries today (see Zelizer 1979 for a discussion of
the role of religion in creating cultural opposition to life insurance). Followers
of Islam have traditionally disapproved of life insurance because it is considered
a hedge against the will of Allah.9 Unsurprisingly, Browne and Kim (1993) and
Meng (1994) find a dummy variable for Islamic countries to be negatively cor-
related with demand for life insurance. Here we use a broader measure of reli-
gious inclination by including Protestantism, Catholicism, and a composite of
other religions, defined as the ratio of the adherents of a religion to the entire
population. While we expect the share of the population that is Muslim to be
7. However, as pointed out by Browne and others (2000), citing unpublished work by Fran�ois
Outreville and George Szpiro, risk aversion might also be negatively correlated with education.
8. We are grateful to one of the referees for pointing this out. A similar debate on the role of educa-
tion has taken place in the empirical growth literature; see Bils and Klenow (2000).
9. The advent of takaful insurance--approved by Islamic scholars and licensed and marketed in
countries with Muslim populations--in the past decade, however, has increased the acceptance of life
insurance in some Islamic populations. For further information see www.insurance.com.my/zone_takaful/
introduction.htm.
Beck and Webb 61
negatively related to demand for life insurance, we do not have prior expecta-
tions about the signs on the other religion variables.
Economies with greater urbanization (a larger share of urban population in
the total) are expected to have higher life insurance consumption. The concen-
tration of consumers in a geographic area simplifies the distribution of life in-
surance products because it reduces the costs related to marketing, premium
collection, underwriting, and claims handling. A larger share of urban popula-
tion is also correlated with less reliance on informal insurance agreements and
therefore may induce higher demand for formal insurance products.
Economic Determinants
Life insurance consumption should rise with income for several reasons. First, a
person's consumption and human capital typically increase along with income,
creating a greater demand for insurance (mortality coverage) to safeguard the
income potential of the insured and the expected consumption of his or her de-
pendents (this would be reflected by a higher TC in equation 1). Second, life
insurance may be a luxury good, since increasing income may enable people to
direct a larger share of their income to retirement and investment-related life
insurance products. Finally, the overhead costs associated with administering
and marketing insurance can make larger policies less expensive per dollar of
insurance in force, lowering their price. Using both aggregate national accounts
data and individual household data, several studies have shown that the use of
life insurance is positively related to income (Campbell 1980; Lewis 1989;
Beenstock and others 1986; Truett and Truett 1990; Browne and Kim 1993;
Outreville 1996). We use real gdp per capita as well as an indicator of perma-
nent income, calculated as the predicted value from a regression of the log of
each country's real gdp per capita on a time trend. Insurance against mortality
risk and consumption and savings decisions are related to permanent income or
income over the life cycle rather than current income.
Theory suggests an ambiguous relationship between life insurance and an
economy's private savings rate. If private agents save a larger share of their in-
come, they might or might not be willing to increase their savings in life insur-
ance policies. We use the share of private savings in gross national disposable
income.
We expect inflation and its volatility to have a negative relationship with life
insurance consumption. Because life insurance savings products typically pro-
vide monetary benefits over the long term, monetary uncertainty has a substan-
tial negative effect on the expected returns on these products. Inflation can also
have a disruptive effect on the life insurance industry when interest rate cycles
spur disintermediation.10 These dynamics make inflation an additional encum-
brance on the product pricing decisions of life insurers, possibly reducing sup-
10. Fixed interest rates and loan options embedded in some life insurance policies, for example,
spurred disintermediation in the U.S. life insurance market during the inflationary 1970s and 1980s.
62 the world bank economic review, vol. 17, no. 1
ply in times of high inflation (see Cargill and Troxel 1979 for a discussion of the
possible effects of inflation on the life insurance market). In addition to testing
for a relationship between life insurance consumption and the inflation rate and
its standard deviation, we also test for a relationship with the real interest rate,
defined as the difference between the nominal interest rate and inflation. Theory
predicts a positive relationship: a higher real interest rate increases life insurers'
investment returns and thus their profitability, in turn offering greater profit-
ability of financial relative to real investments for potential purchasers of life
insurance policies.
We expect banking sector development to be positively correlated with life
insurance consumption.11 Well-functioning banks may increase the confidence
of consumers in other financial institutions, such as life insurers. They also pro-
vide life insurers with an efficient payment system. Efficient development of the
entire financial system--as might be reflected in the absence of interest rate ceil-
ings and other distortionary policies--is thought to help life insurers invest more
efficiently. But a vibrant insurance sector might also foster the development of the
banking sector, so a positive relationship between the two variables cannot neces-
sarily be interpreted as evidence of causality. Outreville (1996) finds a significantly
positive relationship between financial development and life insurance penetra-
tion. We use the total claims of deposit money banks on domestic nonfinancial
sectors as a share of gdp as an indicator of banking sector development.
We expect the size of a country's social security system to be negatively cor-
related with the demand for life insurance products. Kim (1988) and Meng (1994)
postulate that social security displaces private insurance. If greater retirement
savings are being channeled through the government, or if the public sector pro-
vides substantial benefits to families of prematurely deceased wage earners, there
should be less demand for life insurance products (this would be reflected in a
higher W in equation 1). We use public expenditures on social security and welfare
as a share of gdp as an indicator of the size of the social security system.
The correlation of a country's income distribution (as measured by the Gini
index) with life insurance consumption is expected to be ambiguous. Beenstock
and others (1986) reason that wealthy population groups do not need insurance
protection, whereas poorer groups have limited demand because of income con-
straints. (Both the possibility of declining risk aversion with greater wealth and
the replacement of life insurance coverage with surplus assets in an individual's
portfolio are expected to reduce the demand for life insurance among the
wealthy.) A more equal income distribution with a larger middle class might
therefore result in greater demand for life insurance. But although the middle
class may have the greatest demand for life insurance savings products, there
may be a minimum level of income at which these policies become affordable.
Accordingly, in a poor country with a large middle class, fewer people may be
11. Outreville (1992) also proposes a relationship between financial development and insurance
markets.
Beck and Webb 63
able to purchase life insurance than in a poor country with a less equal distribu-
tion and a larger or wealthier upper class. The relationship between income dis-
tribution and life insurance consumption is thus ambiguous. Beenstock and others
(1986) find a negative relationship between the Gini index and life insurance
penetration.
We also test for a relationship between life insurance consumption and the
human development index, as constructed by the United Nations Development
Programme (undp). This index measures the relative achievements of a coun-
try in life expectancy, education (both literacy and gross enrollment), and in-
come (gdp per capita), averaged over the three areas. Values are bounded
between zero and one. Because we expect an ambiguous relationship between
life expectancy and life insurance consumption, we do not necessarily expect a
robust relationship between the human development index and our measures
of life insurance consumption. Outreville (1996) finds no significant relation-
ship between the human development index and life insurance consumption,
and Outreville (1999) shows that the index is positively correlated with mea-
sures of financial development.
Institutional Determinants
A vibrant life insurance market depends to a large extent on the institutional
framework and political stability of a country. If fraud is common in claims re-
porting, insurance becomes prohibitively costly for a large part of the popula-
tion. An inability to appeal the breach of life insurance contracts by insurers
reduces the value of such contracts to consumers and may deter them from com-
mitting large sums of money to these products. Lack of property protection and
contract enforcement hampers life insurers' ability to invest efficiently and con-
trol the price of their products. Finally, lack of political stability shortens the
economic horizon of both potential buyers and suppliers of life insurance prod-
ucts, dampening the development of a healthy life insurance market.
To measure these institutional and political factors, we use three different
indicators. Rule of law measures the degree to which citizens of a country are
able to use the legal system to mediate disputes and enforce contracts. The aver-
age number of revolutions and coups a year indicates the political stability of a
country. Institutional development is an average of six indicators measuring voice
and accountability, political stability, government effectiveness, regulatory qual-
ity, rule of law, and control of corruption. While data for rule of law are avail-
able for 1982�2000 and data for revolutions and coups for 1961�90, data for
institutional development are available for only one point in time, 1998. We there-
fore use this indicator only in the cross-country estimations.
Descriptive Statistics and Correlations
As can be seen in table 1, there is a large variation in the economic and financial
development of countries, their demographic structure, and their macroeconomic
performance. Most of the explanatory variables are correlated with life insur-
64 the world bank economic review, vol. 17, no. 1
ance consumption at the 1 percent level, with the notable exception of the real
interest rate and revolutions and coups (table 2, and appendix tables A-2 and
A-3). But not all the correlations confirm the theoretical predictions. Countries
with a smaller share of young population and a larger share of old population
have higher life insurance consumption, as do those with a longer life expectancy.
Life insurance consumption is also higher for countries in which governments spend
more on transfers and other subsidies and in which income distribution is more
equal.
Many of the potential determinants of life insurance consumption are highly
correlated with one another. Richer countries have older populations, longer
life expectancies, higher levels of schooling, lower inflation, and better devel-
oped banking systems. Countries with higher young dependency ratios have
lower old dependency ratios, shorter life expectancies, and lower levels of edu-
cation. The high correlations between the explanatory variables underscore the
importance of performing multivariate regression analysis as well as the need
to control for country-specific effects that might drive several or all of these
explanatory variables.
III. Empirical Results
Because of the significant correlations between many of the possible determi-
nants of life insurance consumption, we conduct multivariate regression analysis
to assess which determinants robustly predict life insurance consumption even
after we control for other potential effects. The baseline regression includes
real gdp per capita, young and old dependency ratios, average years of school-
ing, life expectancy, the inflation rate, and banking sector development.12 Subse-
quent regressions include a larger set of potential determinants of life insurance
consumption.
Panel Analysis, 1961�2000
Our main results are based on an unbalanced panel of 68 economies, with data
averaged over eight five-year periods (appendix table A-4).13 Using a panel al-
lows us to exploit both cross-country and time-series variation in the data and
to control for differences across countries and over time not accounted for by
any of the explanatory variables.14 We therefore control for both fixed country-
12. We include the dependent and several independent variables in logs so that the coefficients can
be interpreted as elasticities.
13. The number of economies varies across the life insurance measures, and the samples do not overlap
completely.
14. These explanatory variables can be variables that are not included in our estimation because
they do not vary over time or other underlying country characteristics that are not captured in any of
our variables. Among these omitted variables might be the regulation of the insurance sector, taxation,
and the price variable, for which we use proxy variables (such as the supply determinants described in
the section on theoretical underpinnings), but we do not have any direct measures.
Beck and Webb 65
and time-specific effects in our regression and estimate the regression with ei-
ther a fixed or random effects model.15 We average data over five years because
several of our explanatory variables are available only at a five-year frequency
and others might be subject to short-term fluctuations related to the business
cycle.16
The results in table 4 show that the variation in life insurance penetration across
countries can be explained by variation in income, the old dependency ratio,
inflation, and banking sector development. These four variables show signifi-
cant coefficients in our baseline regression and in most of our robustness tests.
Schooling, life expectancy, and the young dependency ratio are not robust pre-
dictors of life insurance consumption.
The results of our baseline regression indicate that a 10 percent increase in
real income per capita increases life insurance penetration by 5.7 percent, con-
firming that life insurance is a luxury good (column 1 of table 4). When we in-
clude the private savings rate and revolutions and coups, however, the coefficient
on income turns insignificant, a result of the smaller sample when either of the
two variables is included.17 When we replace gdp per capita with permanent
income, the results are confirmed (column 9 of table 4).
We find a positive relationship between the old dependency ratio and life in-
surance penetration. The size of the coefficient indicates that a 10 percent in-
crease in the ratio of the old population to the working-age population increases
life insurance penetration by 12 percent. This suggests that demand for savings
and annuity products increases as the population ages.
Price stability is an important predictor of life insurance consumption. The
coefficient on the inflation rate is significantly negative in all specifications. The
effect of a stable macroeconomic environment is also large. If Brazil, which had
one of the highest five-year average inflation rates in our sample, had achieved
an average inflation rate in 1991�95 of 7 percent (the sample median) rather
than the actual 212 percent, life insurance penetration might have been 0.87
percent of gdp rather than 0.29 percent.18 Replacing the inflation rate with the
expected inflation rate--the average of the inflation rate in the current and fol-
15. We test for the appropriateness of the fixed- or random-effects model with the Hausman test.
Under the null hypothesis that random and fixed effects estimates are not statistically different, both
estimators are consistent, but the fixed-effects model is inefficient. Under the alternative hypothesis that
both estimates are statistically different, only the fixed-effects model gives consistent coefficients. We
use the fixed-effects model when the null hypothesis is rejected at the 10 percent level and the random-
effects model otherwise.
16. Average years of schooling are available only at a five-year frequency, and life expectancy, the
urban population share, and the Gini index are not available on a yearly frequency for most countries.
Moreover, the inflation rate and banking sector development might be subject to short-term fluctua-
tions related to the business cycle.
17. We rerun the regressions without the private savings rate or revolutions and coups but restrict-
ing the sample accordingly. In neither case does income per capita enter significantly.
18. This result matches the finding by Babbel (1981) that even the demand for inflation-indexed life
insurance policies decreases during inflationary periods in Brazil.
Table 4. Determinants of Life Insurance Penetration in a Panel, 1961�2000:
Full Sample, Fixed Effects
Variable (1) (2) (3) (4) (5) (6)
Constant �7.069 �8.372 �6.073 �5.662 �8.011 �7.133
(1.92)* (2.07)** (.086) (1.47) (1.06) (1.94)*
GDP per capita 0.567 0.552 0.770 0.424 0.180 0.580
(2.89)*** (2.80)*** (2.82)*** (2.00)** (0.69) (2.96)***
Young dependency �0.357 �0.326 �0.369 �0.964 �0.079 �0.354
ratio (1.12) (1.02) (1.01) (2.73)*** (0.22) (1.11)
Old dependency 1.196 1.195 0.920 1.105 1.308 1.192
ratio (3.90)*** (3.89)*** (2.43)** (2.75)*** (3.82)*** (3.89)***
Life expectancy �0.168 �0.091 �0.900 0.356 0.415 �0.178
(0.22) (0.12) (0.52) (0.46) (0.22) (0.23)
Schooling �0.048 �0.129 0.586 0.194 0.043 �0.054
(0.23) (0.55) (1.85)* (0.87) (0.14) (0.26)
Inflation rate �1.028 �1.038 �1.396 �1.058 �0.827
(5.18)*** (5.22)*** (4.70)*** (5.50)*** (3.37)***
Banking sector 0.352 0.353 0.438 0.227 0.422 0.353
Development (4.62)*** (4.62)*** (4.70) *** (2.80)*** (5.11)*** (4.64)***
Urbanization 0.277
(0.78)
Gini index 0.002
(0.17)
Social security 0.051
(0.50)
Revolutions and �0.065
coups (0.56)
Expected inflation �1.025
rate (5.22)***
Inflation volatility
Real interest rate
Permanent income
Secondary enrollment
Human development
index
Rule of law
Private savings rate
F-test time dummies 2.67** 2.59** 3.23*** 1.31 3.95*** 2.75***
Observations 322 322 177 277 190 322
Economies 66 66 58 61 53 66
Period 1961�2000 1961�2000 1961�2000 1966�2000 1961�90 1961�2000
R2 within 0.6627 0.6635 0.7275 0.6808 0.6387 0.6631
R2 between 0.3234 0.3086 0.4014 0.2889 0.2414 0.3234
R2 overall 0.3827 0.3679 0.5074 0.349 0.3372 0.3822
Hausman test 0.001 0.001 0.001 0.001 0.001 0.001
(p-value)
*Significant at the 10 percent level.
**Significant at the 5 percent level.
***Significant at the 1 percent level.
Note: The numbers in parentheses are t-statistics.
aDeveloping economies, random effects
Source: Authors' calculations.
(7) (8) (9) (10) (11) (12) (13) (14)a
�6.839 �4.578 �9.357 �7.895 �5.136 �7.085 �3.715 �2.380
(1.86)* (1.25) (2.44)** (2.00)** (1.89)* (1.38) (0.64) (0.41)
0.503 0.699 0.668 0.795 �0.017 0.375
* (2.50)** (3.54)*** (3.24)*** (2.73)*** (0.08) (1.72)*
�0.405 �0.518 �0.270 �0.465 �0.681 �0.565 �0.930 �0.042
(1.27) (1.59) (0.84) (1.36) (1.65) (1.04) (2.47)** (0.06)
1.159 1.230 1.076 1.238 1.757 1.471 1.302 0.226
* (3.77)*** (3.61)*** (3.42)*** (3.65)*** (3.96)*** (2.23)** (3.24)*** (0.40)
�0.098 �0.779 �0.069 �0.115 �0.470 0.212 �0.963
(0.13) (1.01) (0.09) (0.15) (0.51) (0.16) (0.94)
�0.029 �0.221 �0.097 �0.294 0.420 �0.075
(0.14) (1.04) (0.47) (0.88) (1.64) (0.23)
�0.806 �1.659 �1.060 �1.047 �1.049 �0.984 �1.138 �1.187
(3.15)*** (5.79)*** (5.36)*** (5.26)*** (4.85)*** (4.27)*** (4.98)*** (4.28)***
0.354 0.331 0.349 0.332 0.344 0.268 0.368 0.598
* (4.65)*** (4.14)*** (4.62)*** (4.09)*** (3.76)*** (2.38)** (4.69)*** (3.76)***
*
�0.056
(1.37)
0.302
(2.78)***
0.792
(3.28)***
�0.029
(0.14)
2.423
(1.51)
�0.001
(0.02)
0.359
(2.56)**
* 1.89* 2.45** 2.02* 1.76 0.46 1.12 3.39** 4.02
2 322 304 322 298 266 224 205 141
6 66 64 66 65 65 66 57 37
0 1961�2000 1961�2000 1961�2000 1961�2000 1976�2000 1981�2000 1971�95 1961�2000
1 0.6653 0.6895 0.6659 0.6573 0.6073 0.5723 0.7189 0.4856
4 0.3266 0.3235 0.3328 0.3252 0.3 0.3749 0.3791 0.1219
2 0.3876 0.366 0.3834 0.3889 0.3428 0.3695 0.3957 0.1777
0.001 0.001 0.001 0.001 0.058 0.001 0.001 0.444
68 the world bank economic review, vol. 17, no. 1
lowing year--confirms the results (column 6 of table 4).19 Inflation volatility does
not explain any variation in life insurance penetration across countries, whereas
the real interest rate is positively related to life insurance penetration when in-
flation is controlled for (columns 7 and 8 of table 4).
Banking sector development is positively correlated with life insurance pen-
etration. The coefficient on the indicator of banking sector development is sig-
nificantly positive in all specifications. As discussed, the positive coefficient does
not imply a causal effect on life insurance penetration. Instead, it shows that
countries that have well-developed banks also have higher life insurance con-
sumption. In our cross-country analysis we try to control for reverse causation
and simultaneity bias.
Variation in the share of young population or in life expectancy cannot ex-
plain the variation in life insurance penetration across countries, confirming the
hypothesis of offsetting effects of the young dependency ratio (life expectancy)
on gross premiums, a positive (negative) effect on mortality risk, and a negative
(positive) effect on the savings and annuity components.20 Neither average years
of schooling nor secondary enrollment enter significantly at the 5 percent level
in any of the regressions.
Turning to our additional explanatory variables, we find a positive relation-
ship between the private savings rate and life insurance penetration. Urbaniza-
tion (column 2 of table 4), the Gini index (column 3), social security (column 4),
revolutions and coups (column 5), the human development index (column 11),
and rule of law (column 12) cannot explain the cross-country variation in life
insurance penetration.21 In the baseline regression with the sample limited to
developing economies, only inflation and banking sector development continue
to enter significantly at the 1 percent level, whereas income per capita enters
significantly and positively at the 10 percent level (column 14 of table 4). The
old dependency ratio cannot explain the variation in life insurance penetration
across developing economies.
Table 5 presents results with the other indicators of life insurance consump-
tion across countries as dependent variables. For each indicator it gives results
for two baseline regressions, one for the full sample and one restricted to de-
veloping economies. Life insurance density increases with higher income per
capita, a higher old dependency ratio, a lower inflation rate, and better devel-
oped banks (column 1 of table 5). Once we restrict the sample to developing
19. Following Browne and Kim (1993), we also use the average of inflation in the current and pre-
vious year, because consumers' inflation expectations might be determined by previous inflation expe-
rience. The results do not change.
20. Because the young and old dependency ratios and life expectancy are highly correlated with one
another, this result might be driven by multicollinearity. We therefore test the robustness of the results
by including only one of the three variables at a time. The results do not change.
21. We also try two alternative indicators of institutional development, corruption and bureaucratic
quality (like rule of law, these indicators come from the Political Risk Services (various years) Interna-
tional Country Risk Guide. Neither enters significantly in the regressions.
Beck and Webb 69
economies, however, only banking sector development enters significantly. The
income elasticity of life insurance density is higher than that of life insurance
penetration, as expected (see the discussion in the section on measuring life
insurance consumption).
Life insurance in private savings increases with a higher old dependency ratio,
lower inflation, and better developed banks (column 3 of table 5). Interestingly,
the share of savings in life insurance policies decreases with a higher savings rate.
Considering this result jointly with the positive coefficient (0.359) on the sav-
ings rate in the regression of life insurance penetration (column 13 of table 4)
suggests that although private agents invest some of their additional savings in
life insurance policies, overall there is a shift in their portfolios away from life
insurance policies toward other savings instruments. gdp per capita does not
explain the share of savings in life insurance policies. In the sample of develop-
ing economies only banking sector development (positively) and the private sav-
ings rate (negatively) can explain the variation in the share of private savings in
life insurance policies across developing economies.
Life insurance in force to gdp increases with higher income per capita, lower
inflation, a lower old dependency ratio, and better developed banks. While the
results for gdp per capita, inflation, and banking sector development confirm
the results using life insurance penetration and life insurance density, the results
for the old dependency ratio are surprising. The stronger weight of the mortal-
ity risk component in life insurance in force to gdp compared with that in the
other three measures, and its exclusion of annuities, might explain the opposite
sign.22 Only the results for income per capita and inflation are confirmed in the
sample restricted to developing economies.
Annual Panel, 1961�2000
Table 6 presents results for a panel of annual observations. Using annual rather
than five-year averages allows us to maximize the information we have and to
test the sensitivity of our panel analysis to the frequency of the data.23 As in the
five-year panel, life insurance penetration increases with income per capita, the
old dependency ratio, and banking sector development and decreases with in-
flation. Interestingly, we also find a negative relationship between the young
dependency ratio and life insurance penetration, suggesting that countries with
a larger share of young population have lower life insurance consumption.24 As
in the five-year panel, expected inflation has a negative relationship with life
insurance penetration (column 3 of table 6), and the real interest rate, perma-
22. This might also explain the negative sign on life expectancy. In regressions with only the old or
the young dependency ratio or life expectancy, only the old dependency ratio and life expectancy enter
negatively and significantly at the 5 percent level.
23. Because schooling data are available only at a five-year frequency, we repeat the values for the
intermediate years from the initial year of the corresponding five-year period.
24. As in the five-year panel, we include the young and old dependency ratios and life expectancy
separately, confirming our results.
Table 5. Determinants of Life Insurance Consumption in a Panel, 1961�2000, with Alternative
Measures of Life Insurance Consumption
70 (1) (2) (3) (4) (5) (6)
Life insurance density Life insurance in private savings Life insurance in force to gdp
Developing Developing Developing
Full sample, economies, Full sample, economies, Full sample, economies,
Variable fixed effects fixed effects fixed effects random effects random effects random effects
Constant �13.342 �19.270 1.232 �4.909 8.977 7.699
(2.13)** (1.69)* (0.18) (0.46) (1.21) (0.63)
gdp per capita 1.471 0.745 �0.254 �0.432 0.924 0.759
(4.41)*** (1.09) (0.98) (1.23) (3.81)*** (2.28)**
Young dependency ratio �0.299 1.208 �0.756 1.000 0.258 �0.428
(0.55) (0.82) (1.71)* (1.03) (0.47) (0.38)
Old dependency ratio 1.730 0.885 1.604 1.511 �1.313 �1.423
(3.31)*** (0.41) (3.40)*** (1.41) (3.03)*** (1.40)
Life expectancy 0.023 1.392 0.188 0.111 �3.403 �1.644
(0.02) (0.69) (0.12) (0.05) (1.94)* (0.51)
Schooling �0.169 �0.054 0.586 0.038 0.572 �0.231
(0.48) (0.06) (1.95)* (0.08) (1.39) (0.41)
Inflation rate �0.757 �0.600 �0.706 �0.473 �1.394 �1.979
(2.24)** (1.15) (2.62)*** (1.18) (2.64)*** (2.25)**
Banking sector 0.375 0.938 0.371 0.750 0.446 0.204
development (2.89)*** (2.69)*** (4.02)*** (3.22)*** (2.91)*** (0.58)
Private savings rate �0.660 �0.561
(3.97)*** (2.30)**
F-test time dummies 1.17 0.18 3.40** 8.33* 15.05** 6.87
Observations 322 141 203 88 216 75
Economies 66 37 56 28 47 22
Period 1961�2000 1961�2000 1971�95 1971�95 1961�95 1961�95
R2 within 0.6057 0.2437 0.7002 0.5517 0.3525 0.4181
R2 between 0.7146 0.3278 0.1914 0.0228 0.4895 0.3906
R2 overall 0.7211 0.3141 0.2878 0.0756 0.4256 0.3437
Hausman test (p-value) 0.001 0.021 0.0092 0.4985 0.491 0.9615
71 *Significant at the 10 percent level.
**Significant at the 5 percent level.
***Significant at the 1 percent level.
Note: The numbers in parentheses are t-statistics.
Source: Authors' calculations.
Table 6. Determinants of Life Insurance Penetration in an Annual Panel, 1961�2000
(1) (2) (3) (4) (5) (6)
72
Full Sample Developing
economies
Variable Fixed effects Random effects Fixed effects Fixed effects Fixed effects Random effects
Constant �3.288 13.928 �5.831 �1.828 �6.478 �3.694
(1.16) (3.01)*** (2.19)** (0.62) (2.21)** (0.80)
gdp per capita 0.665 0.088 0.517 0.689 0.394
(4.89)*** (0.72) (3.92)*** (4.96)*** (2.15)**
Young dependency ratio �0.586 �0.712 �0.401 �0.713 �0.474 0.069
(2.87)*** (2.94)*** (2.09)** (3.32)*** (2.32)** (0.13)
Old dependency ratio 0.920 0.506 1.137 0.901 0.689 0.121
(4.97)*** (2.40)** (6.55)*** (4.38)*** (3.53)*** (0.25)
Life expectancy �0.631 �4.056 �0.406 �0.890 �0.601 �0.688
(1.02) (3.59)*** (0.68) (1.37) (0.99) (0.81)
Schooling �0.068 0.852 0.010 �0.156 �0.141 �0.154
(0.45) (4.49)*** (0.07) (0.97) (0.94) (0.56)
Inflation rate �0.645 �0.687 �0.788 �0.686 �0.708
(4.97)*** (4.38)*** (4.41)*** (5.32)*** (3.88)***
Banking sector 0.062 0.412 0.405 0.083 0.060 0.727
development (2.29)** (8.47)*** (9.55)*** (2.74)*** (2.20)** (6.87)***
Private savings rate 0.184
(2.47)**
Expected inflation rate �0.834
(6.13)***
Real interest rate 0.172
(2.07)**
Permanent income 1.034
(5.91)***
F-test time dummies 2.03*** 80.21*** 2.01*** 1.73*** 1.47** 20.34
Observations 836 463 779 782 836 288
Economies 66 55 66 63 66 37
Period 1961�2000 1970�95 1961�2000 1961�2000 1961�2000 1961�2000
R2 within 0.6166 0.6716 0.6589 0.6151 0.6221 0.4383
73 R2 between 0.3106 0.4882 0.3396 0.3072 0.3292 0.1289
R2 overall 0.3767 0.5245 0.4027 0.3523 0.3916 0.1945
Hausman test (p-value) 0.001 0.1384 0.001 0.001 0.001 0.9758
*Significant at the 10 percent level.
**Significant at the 5 percent level.
***Significant at the 1 percent level.
Note: The numbers in parentheses are t-statistics.
Source: Authors' calculations.
74 the world bank economic review, vol. 17, no. 1
nent income, and the private savings rate enter positively (columns 2, 4, and 5).
Neither schooling nor life expectancy shows a robust relationship with life in-
surance penetration. Only income per capita, inflation, and banking sector de-
velopment explain the variation in life insurance penetration across developing
economies in the annual sample. Overall, the annual sample thus confirms the
findings from the five-year panel regressions.
Cross-Country Analysis, 1980�2000
Table 7 presents results from cross-country regressions in which we average data
over the period 1980�2000 for all economies in our sample. Although cross-
country analysis does not allow us to control for omitted variables, as in the panel
analysis, it does permit us to test the relationship between life insurance con-
sumption across countries and several time-invariant variables and to use instru-
mental variables regressions to control for biases induced by simultaneity and
reverse causation. These biases might arise especially for educational attainment
and banking sector development.
Countries with higher levels of economic and financial development, a more
educated population, lower inflation, and a shorter life expectancy have higher
life insurance penetration. Moreover, the old dependency ratio enters negatively
and significantly at the 10 percent level.25 While the results for income per capita,
inflation, and banking sector development confirm the results from our panel
analysis, those for life expectancy, schooling, and the old dependency ratio dif-
fer from the previous results. Restricting the sample to developing economies
confirms the results for life expectancy, inflation, schooling, and the old depen-
dency ratio but not for income per capita and banking sector development. The
young dependency ratio, the private savings rate, and revolutions and coups do
not enter significantly in the regressions (columns 3 and 5 of table 7). A larger
share of Muslim population reduces life insurance penetration, and a better in-
stitutional environment increases it (columns 4 and 6 of table 7).
Econometric, sampling, and frequency differences might explain the differ-
ences between the panel and cross-country results. The panel estimations allow
us to control for country-specific effects, whereas the ordinary least squares re-
gressions do not.26 Moreover, economic and demographic factors might have
different relationships with life insurance consumption across countries than
within countries over time.
Our cross-country results show a positive relationship between schooling and
banking sector development and life insurance consumption. But these results
25. As in the five-year panel, we control for multicollinearity by including only one of the following
regressors at a time: the old dependency ratio, the young dependency ratio, and life expectancy. Al-
though life expectancy continues to enter significantly and negatively, neither of the two dependency
ratios enters significantly.
26. Most developing economies do not have life insurance data for the period before 1978, so the
unbalanced panel regressions might be biased toward developed countries. We therefore rerun all regres-
sions of the five-year panel with the sample limited to 1981�2000. The results do not change significantly.
Beck and Webb 75
do not allow any inferences about a causal relationship between education and
banking, on the one hand, and the development of the life insurance sector on
the other. We therefore run two instrumental variables regressions in which we
extract the exogenous components of banking sector development and school-
ing to control for reverse causation and simultaneity bias in the empirical rela-
tionship between these variables and life insurance consumption. Specifically,
we use dummy variables indicating the origin of a country's legal system and a
variable--good crops--proxying for agricultural endowments conducive to a
large middle class and institutional development.27 Legal origin and agricultural
endowments are both exogenous variables and are highly correlated with bank-
ing sector development and schooling, as confirmed by the first-stage regressions
(the two variables explain 43 percent of the variation in banking sector develop-
ment and schooling). We use the Hansen test of overidentifying restrictions to
examine whether legal origin and agricultural endowments have any effect on
life insurance penetration beyond their effect through banking sector develop-
ment, schooling, or the other explanatory variables. In column 7 of table 7 we
instrument only for banking sector development, and in column 8 for both bank-
ing sector development and schooling.
While banking sector development enters significantly and positively even after
we instrument for it, schooling turns insignificant when we instrument for it.
The test of overidentifying restrictions is not rejected in either case, confirming
the adequacy of our instruments.28 These results show that the relationship be-
tween banking sector development and life insurance consumption is not due to
reverse causation and simultaneity bias, and the significant relationship between
schooling and life insurance consumption is most likely spurious.
Overall, the cross-country results confirm the importance of income per capita,
monetary stability, and banking sector development in predicting life insurance
consumption across countries. They also provide evidence of the importance of
religion and institutional development for life insurance consumption. Finally,
the demographic variables show a different relationship with life insurance con-
sumption in the cross-section than in the panel.
IV. Conclusion
In this article we analyze the determinants of life insurance consumption in a panel
of 68 economies for 1961�2000, using four different indicators of life insurance
consumption. Our main results are based on a panel of eight nonoverlapping five-
year periods. We test for the sensitivity of the results with a panel of annual obser-
vations and a cross-country sample.
27. Beck and others (forthcoming), among many others, show that legal origin explains the varia-
tion in financial development across countries. Easterly and Levine (2003) show that good crops are a
good predictor of institutional development.
28. We also ran an instrumental variables regression in which we instrumented only schooling. The
test of overidentifying restrictions is rejected, however.
Table 7. Determinants of Life Insurance Penetration in a Cross-Section, 1980�2000
(1) (2) (3) (4) (5) (6) (7) (8)
Developing Full Sample
Full sample, economies,
Variable ols ols ols ols ols ols iv iv
76
Constant 39.729 51.455 32.775 32.466 41.322 41.991 33.825 29.156
(2.98)*** (2.24)** (3.08)*** (2.83)*** (2.97)*** (3.17)*** (2.74)*** (1.97)*
gdp per capita 0.595 0.621 0.342 0.660 0.616 0.396 0.560 0.628
(3.45)*** (1.68) (1.63) (4.48)*** (3.52)*** (2.63)** (2.76)*** (2.62)**
Young dependency ratio �0.980 �2.237 �0.397 �0.274 �1.349 �0.806 �0.190 �0.192
(1.00) (1.00) (0.41) (0.39) (1.19) (0.85) (0.17) (0.16)
Old dependency ratio �0.665 �1.487 0.188 �0.374 �0.775 �0.759 �0.734 �0.711
(1.68)* (1.79)* (0.39) (1.00) (1.84)* (1.84)* (1.10) (0.96)
Life expectancy �10.618 �11.576 �10.113 �9.543 �10.599 �10.853 �10.961 �10.067
(4.40)*** (3.51)*** (5.68)*** (3.97)*** (4.36)*** (4.61)*** (4.48)*** (3.45)***
Schooling 1.824 1.904 2.118 1.137 1.705 1.554 1.871 1.021
(5.06)*** (2.92)*** (4.98)*** (2.35)** (4.45)*** (3.97)*** (3.74)*** (0.89)
Inflation rate �1.830 �2.003 �2.183 �2.102 �1.756 �1.213 �0.371 0.286
(4.10)*** (2.11)** (2.73)*** (5.06)*** (3.39)*** (2.38)** (0.32) (0.19)
Banking sector 0.631 0.298 0.234 0.750 0.622 0.639 1.802 2.229
development (2.59)** (0.69) (0.60) (3.67)*** (2.44)** (2.51)** (2.44)** (2.32)**
Private savings rate 1.284
(1.37)
Muslim �0.018
(2.60)**
Catholic �0.002
(0.59)
Protestant �0.009
(1.42)
Revolutions and coups 0.508
(1.39)
Institutional development 0.729
(2.22)**
Observations 66 37 58 66 63 66 62 62
R2 0.70 0.54 0.74 0.78 0.70 0.72 0.61 0.51
Instrumented Banking sector Banking sector
development development,
schooling
Instruments Legal origin Legal origin
77 dummy variables, dummy variables,
good crops good crops
Test of overidentifying 0.1793 0.2374
restrictions (p-value)
ols ordinary least squares.
iv instrumental variables.
*Significant at the 10 percent level.
**Significant at the 5 percent level.
***Significant at the 1 percent level.
Note: The numbers in parentheses are t-statistics.
Source: Authors' calculations.
78 the world bank economic review, vol. 17, no. 1
Our panel estimations show that countries with higher income (both cur-
rent and permanent), lower inflation, and better developed banks have higher
life insurance consumption. A higher ratio of old to working-age population
increases life insurance penetration and life insurance density, and it decreases
life insurance in force to gdp, perhaps reflecting the different weights of mor-
tality risk, savings, and annuity components in these measures. A higher pri-
vate savings rate and a higher real interest rate are also associated with higher
life insurance consumption. The young dependency ratio, life expectancy, and
schooling have no strong association with life insurance consumption across
countries.
The share of life insurance premiums in private savings is best predicted by
the old dependency ratio, inflation, banking sector development, and the pri-
vate savings rate but not by income per capita. The results suggest that the older
the population and the lower the inflation rate, the more people will select life
insurance over other forms of savings. But as private agents save more, the share
of life insurance in their portfolios declines even though they invest some of their
additional savings in life insurance policies. Restricting the sample to develop-
ing economies makes many of the results less significant, but macroeconomic
stability and well-developed banks continue to predict higher life insurance con-
sumption across developing economies.
The cross-country estimations confirm some of the panel results and con-
tradict others. Most notably, we find a positive relationship between school-
ing and life insurance consumption, though it is not robust to controlling for
biases induced by reverse causation and simultaneity. By contrast, the positive
effect of banking sector development on life insurance consumption is robust
to controlling for these biases by instrumenting with legal origin and agricul-
tural endowments. This evidence suggests that banking sector development
facilitates the development of life insurance and its contractual savings func-
tion. This finding does not contradict the positive effect of life insurance on
capital market development found by other authors. While an efficient bank-
ing system might help develop the life insurance sector by offering payment
services and raising confidence in financial institutions, life insurance and other
forms of contractual savings might foster the development of capital markets
through demand for long-term financial investments.
In summary, income per capita, inflation, and banking sector development
are the most robust predictors of life insurance consumption across countries
and over time. In addition, religious and institutional differences can explain some
of the variation in life insurance consumption across countries. But there is no
robust link from schooling and the demographic variables to life insurance con-
sumption. Finally, although life insurance is a luxury good, there is no relation-
ship between income distribution and life insurance consumption. Rising income
per capita helps drive life insurance consumption, but income distribution does
not appear to do so.
Beck and Webb 79
The results provide a thorough review of existing hypotheses about the de-
mand for and supply of life insurance. They also have implications for policy-
makers. Both monetary stability and banking sector development have positive
effects on economic development and growth independent of their positive ef-
fect on the development of the insurance sector. Moreover, they may be funda-
mental to the growth of savings and investment through life insurance,
particularly in a developing economy.
Appendix Table A-1. Definitions and Sources of Variables
Variable Definition Source
Life insurance penetration Life insurance premiums divided by gdp. Swiss Reinsurance Company (various years);
imf (various years)
Life insurance density Life insurance premiums per capita in real dollars. Calculated as life Swiss Reinsurance Company (various years);
insurance premiums multiplied by the average period exchange imf (various years); World Bank (various years)
rate, divided by the population and the U.S. consumer price index.
Life insurance in private savings Life insurance premiums divided by private savings. Swiss Reinsurance Company (various years);
Loayza and others (1999)
Life insurance in force to gdp Outstanding life insurance policies relative to gdp. Calculated as the American Council of Life Insurance (various years);
sum of face amounts plus dividend additions of life insurance imf (various years)
policies outstanding as a share of gdp.
gdp per capita gdp per capita in constant 1995 U.S. dollars. World Bank (various years)
Young dependency ratio Ratio of the population under age 15 to the population ages 15�65. World Bank (various years)
Old dependency ratio Ratio of the population over age 65 to the population ages 15�65. World Bank (various years)
Life expectancy Years of life expectancy at birth. World Bank (various years)
80
Schooling Average years of schooling in the population over age 25. Barro and Lee (1996, 2000)
Inflation rate Log difference of the consumer price index (line 64 in imf, imf (various years)
International Financial Statistics, various years)
Banking sector development {(0.5) * [F(t)/P_e(t) + F(t � 1)/P_e(t � 1)]}/[GDP(t)/P_a(t)], where F is imf (various years)
claims by deposit money banks and other financial institutions on
domestic nonfinancial sectors (lines 22a�d), gdp is line 99b, P_e is
the end-of-period consumer price index (line 64), and P_a is the
average consumer price index for the year.
Urbanization Share of the urban population in the total population. World Bank (various years)
Gini index The Gini index measures the area between the Lorenz curve (the Deininger and Squire (1996); Lundberg and
cumulative percentages of total income received against the Squire (2001)
cumulative number of recipients) and a hypothetical line of abso-
lute equality, expressed as a percentage of the maximum area under
the line. Thus a Gini index of 0 represents perfect equality, and an
index of 100 perfect inequality.
Social security Government subsidies and other current transfers by government as a World Bank (various years)
share of gdp.
Real interest rate Nominal interest rate minus the inflation rate. The nominal rate is imf (various years)
the average lending rate or, if unavailable, the discount rate.
Expected inflation rate Average of the inflation rate in the current and following year. imf (various years)
Permanent income Predicted value of a regression of the log of each country's real gdp World Bank (various years); authors' calculations
per capita on a time trend.
Secondary enrollment Gross secondary enrollment ratio. World Bank (various years)
Private savings rate Private savings as a share of gross national disposable income. Loayza and others (1999)
Revolutions and coups Average number of revolutions and coups a year. Banks (1994)
Human development index Average of a country's achievements in life expectancy, education undp (2002)
(literacy and gross enrollment), and gdp per capita, normalized
between 0 and 1.
Rule of law Measure of the extent to which the citizens of a country trust the Political Risk Services (various years)
legal system to settle disputes. Values range from 6 (strong law and
order tradition) to 1 (weak law and order tradition).
Inflation volatility Standard deviation of inflation. IMF (various years)
Institutional development Average of six indicators measuring voice and accountability, political
stability, government effectiveness, regulatory quality, rule of law,
and control of corruption. Each of these indicators is constructed
from a wide array of survey indicators. Kaufmann and others (1999)
81 Catholic Share of Catholic adherents in the total population. La Porta and others (1999)
Muslim Share of Muslim adherents in the total population. La Porta and others (1999)
Protestant Share of Protestant adherents in the total population. La Porta and others (1999)
British legal origin Dummy variable that takes the value of 1 if the country's legal system
is of British origin. La Porta and others (1999)
French legal origin Dummy variable that takes the value 1 if the country's legal system
is of French origin. La Porta and others (1999)
Socialist legal origin Dummy variable that takes the value 1 if the country's legal system
is of socialist origin. La Porta and others (1999)
German legal origin Dummy variable that takes the value 1 if the country's legal system
is of German origin. La Porta and others (1999)
Scandinavian legal origin Dummy variable that takes the value 1 if the country's legal system
is of Scandinavian origin. La Porta and others (1999)
Good crops (1 + zmaize + zwheat)/(1 + zrice + zsugarcane), where zX equals the
share of the land area judged by the Food and Agriculture Organi-
zation to be suitable for growing crop X. Maize and wheat are con-
sidered to be crops that foster a large middle class with egalitarian
institutions, while rice and sugarcane tend to produce a powerful
elite and more closed institutions. Easterly and Levine (2003)
82 the world bank economic review, vol. 17, no. 1
Appendix Table A-2. Additional Correlations
Life insurance
Variable penetration Gini index Urbanization Social security
Gini index �0.2626*** 1.0000
Urbanization 0.2390*** �0.3428*** 1.0000
Social security 0.2883*** �0.6471*** 0.5010*** 1.0000
Real interest rate �0.0851 0.2067*** 0.0842* �0.0220
Expected inflation rate �0.2586*** 0.2571*** 0.0917* �0.0739
Permanent income 0.5321*** �0.5737*** 0.5831*** 0.6096***
Secondary enrollment 0.5475*** �0.6193*** 0.6562*** 0.7204***
Private savings rate 0.1902*** �0.2494*** 0.1668*** 0.1749***
Revolutions and coups �0.0698 0.1528** �0.2570*** �0.2867***
Human development index 0.4415*** �0.4772*** 0.7555*** 0.6363***
Rule of law 0.3519*** �0.6012*** 0.5253*** 0.6120***
Inflation volatility �0.2125*** 0.2094*** 0.0385 �0.0700
Institutional development 0.5232*** �0.6076*** 0.6233*** 0.6477***
Catholic �0.2069* 0.2286* 0.0851 0.0348
Muslim �0.3057** 0.0878 �0.2667** �0.2842**
Protestant 0.3165** �0.2236* 0.2488** 0.3201***
British legal origin 0.3364*** 0.1623 �0.1080 �0.1716
French legal origin �0.4565*** 0.4044*** �0.0302 �0.2524**
Socialist legal origin �0.2333* �0.4044*** �0.1189 0.3565***
German legal origin 0.4518*** �0.2774** 0.1484 0.0397
Scandinavian legal origin 0.1000 �0.2591** 0.2296* 0.3059**
Good crops 0.1048 �0.6832*** 0.3525*** 0.5585***
*Significant at the 10 percent level.
**Significant at the 5 percent level.
***Significant at the 1 percent level.
Source: Authors' calculations.
Beck and Webb 83
Real Human
interest Expected Permanent Secondary Private Revolutions development
rate inflation rate income enrollment savings rate and coups index
1.0000
0.6610*** 1.0000
�0.0421 �0.2065*** 1.0000
�0.0461 �0.1225** 0.7617*** 1.0000
0.0799 �0.0136 0.3366*** 0.3194*** 1.0000
0.0756 0.0879 �0.3100*** �0.2630*** �0.1847*** 1.0000
0.0008 �0.0826 0.7705*** 0.8544*** 0.2761*** �0.2812*** 1.0000
�0.0593 �0.2483*** 0.7354*** 0.7050*** 0.3078*** �0.4591*** 0.7611***
0.5876*** 0.9002*** �0.1860*** �0.1120** �0.0398 0.1246** �0.0535
�0.0813 �0.3256*** 0.8112*** 0.8363*** 0.3307** �0.2650** 0.8547***
0.2258* 0.2500** �0.1082 �0.0890 �0.2702** 0.1945 0.1090
�0.0878 �0.0604 �0.3167*** �0.3291*** �0.1323 �0.0248 �0.4936***
�0.0950 �0.2097* 0.5781*** 0.4440*** �0.0659 �0.1618 0.3531***
�0.1192 �0.1887 �0.013 �0.0667 0.2068 0.0077 �0.1031
0.1937 0.1773 �0.3384*** �0.3296*** �0.4266*** 0.1928 �0.2571***
�0.0410 0.2621** �0.2169* 0.0950 0.3284** �0.0810 0.0844
�0.0575 �0.1917 0.4395*** 0.2421** 0.3667*** �0.1227 0.2566**
�0.0553 �0.1303 0.4621*** 0.3842*** �0.1257 �0.1552 0.3070***
�0.0373 0.0841 0.3076** 0.4809*** 0.1646 �0.2639** 0.4157***
Appendix Table A-3. Additional Correlations
British French Socialist German Scandinavian
Rule of Inflation Institutional legal legal legal legal legal
Variable law volatility development Catholic Muslim Protestant origin origin origin origin origin
Inflation volatility �0.2076*** 1.0000
Institutional 0.8513*** �0.2305* 1.0000
development
Catholic �0.1380 0.2596** �0.0200 1.0000
Muslim �0.3333*** �0.1232 �0.4567*** �0.4221*** 1.0000
Protestant 0.4883*** �0.1825 0.4861*** �0.3243*** �0.2348* 1.0000
British legal origin 0.0492 �0.1794 0.1092 �0.3584*** �0.0229 0.0441 1.0000
French legal origin �0.4847*** 0.1201 �0.4114*** 0.5157*** 0.2495** �0.4050*** �0.5568*** 1.0000
Socialist legal origin 0.1489 0.2712 �0.0528 0.0009 �0.1357 �0.1563 �0.2232* �0.3271*** 1.0000
German legal origin 0.2573** �0.1469 0.2632** �0.0812 �0.1356 0.0470 �0.1902 �0.2787** �0.1118 1.0000
Scandinavian legal 0.3785*** �0.0967 0.3804** �0.2847** �0.1244 0.8429*** �0.1723 �0.2525** �0.1012 �0.0863 1.0000
origin
Good crops 0.4268*** 0.0872 0.3737** �0.0166 �0.1789 0.0545 �0.1183 �0.2963** 0.4418*** 0.1977 0.0042
*Significant at the 10 percent level.
**Significant at the 5 percent level.
***Significant at the 1 percent level.
Source: Authors' calculations.
Beck and Webb 85
Appendix Table A-4. Economies in the Sample for Each Measure of Life
Insurance Consumption
Life insurance Life insurance Life insurance
Economy penetration and density in private savings in force to gdp
Algeria *
Argentina * *
Australia * * *
Austria * * *
Belgium * * *
Brazil * * *
Bulgaria *
Cameroon * *
Canada * * *
Chile * * *
China * *
Colombia * *
Costa Rica * * *
Croatia *
Cyprus * *
Czech Republic *
Denmark * * *
Dominican Republic * * *
Ecuador * * *
Egypt, Arab Rep. * * *
El Salvador *
Fiji *
Finland * * *
France * * *
Germany * * *
Great Britain * * *
Greece * * *
Guatemala * * *
Honduras *
Hong Kong, China * *
Hungary *
Iceland * * *
India * * *
Indonesia * * *
Iran, Islamic Rep. of * *
Ireland * * *
Israel * * *
Italy * * *
Japan * * *
Kenya * *
Korea, Rep. of * * *
Malaysia * * *
Mexico * * *
Netherlands * * *
New Zealand * * *
Norway * * *
Pakistan * * *
Panama * *
(continued)
86 the world bank economic review, vol. 17, no. 1
Appendix Table A-4. (continued)
Life insurance Life insurance Life insurance
Economy penetration and density in private savings in force to gdp
Peru * * *
Philippines * * *
Poland * *
Portugal * * *
Romania *
Singapore * *
Slovenia *
South Africa * * *
Spain * * *
Sweden * * *
Switzerland * * *
Syrian Arab Republic *
Taiwan, China * * *
Thailand * * *
Tunisia * * *
Turkey * *
Uruguay * * *
United States * * *
Venezuela, RB * * *
Zimbabwe * *
References
American Council of Life Insurance. Various years. Life Insurance Fact Book. Washing-
ton, D.C.
Babbel, David F. 1981. "Inflation, Indexation and Life Insurance Sales in Brazil." Jour-
nal of Risk and Insurance 48:115�35.
Banks, Arthur S. 1994. Cross-National Time Series Data Archive. Binghamton: State
University of New York at Binghamton, Center for Social Analysis.
Barro, Robert. J., and Jong Wha Lee. 1996. "International Measures of Schooling Years
and Schooling Quality." American Economic Review Papers and Proceedings
86:218�23.
------. 2000. "International Data on Educational Attainment: Updates and Implica-
tions." cid Working Paper 42, Harvard University, Boston.
Beck, Thorsten, Asli Demirg��-Kunt, and Ross Levine. Forthcoming. "Law, Endowments,
and Finance." Journal of Financial Economics.
Beenstock, Michael, Gerry Dickinson, and Sajay Khajuria. 1986. "The Determination
of Life Premiums: An International Cross-Section Analysis, 1970�1981." Insurance:
Mathematics and Economics 5:261�70.
Bils, Mark, and Peter J. Klenow. 2000. "Does Schooling Cause Growth?" American
Economic Review 90:1160�83.
Browne, Mark J., and Kihong Kim. 1993. "An International Analysis of Life Insurance
Demand." Journal of Risk and Insurance 60:616�34.
Beck and Webb 87
Browne, Mark J., Jae Wook Chung, and Edward W. Frees. 2000. "International Prop-
erty-Liability Insurance Consumption." Journal of Risk and Insurance 67:73�90.
Campbell, Ritchie A. 1980. "The Demand for Life Insurance: An Application of the
Economics of Uncertainty." Journal of Finance 35:1155�72.
Cargill, Thomas F., and T. E. Troxel. 1979. "Modeling Life Insurance Savings: Some
Methodological Issues." Journal of Risk and Insurance 46:391�410.
Catalan, Mario, Gregorio Impavido, and Alberto Musalem. 2000. "Contractual Savings
or Stock Market Development: Which Leads?" Policy Research Working Paper 2421.
World Bank, Financial Sector Development Department, Washington, D.C.
Deininger, Klaus, and Lyn Squire. 1996. "A New Data Set Measuring Income Inequal-
ity." World Bank Economic Review 10:565�91.
Easterly, William, and Ross Levine. 2003. "Tropics, Germs and Crops: How Endow-
ments Influence Economic Development." Journal of Monetary Economics 50:3�39.
Hakansson, Nils H. 1969. "Optimal Investment and Consumption Strategies under
Risk, and under Uncertain Lifetime and Insurance." International Economic Review
10:443�66.
imf (International Monetary Fund). Various years. International Financial Statistics.
Washington, D.C.
Impavido, Gregorio, and Alberto Musalem. 2000. "Contractual Savings, Stock, and Asset
Markets." Policy Research Working Paper 2490. World Bank, Financial Sector De-
velopment Department, Washington, D.C.
Kaufmann, Daniel, Aart Kraay, and Pablo Zoido-Lobat�n. 1999. "Aggregating Gover-
nance Indicators." Policy Research Working Paper 2195. World Bank, Development
Research Group, Macroeconomics and Growth; and World Bank Institute, Gover-
nance, Regulation, and Finance, Washington, D.C.
Kim, Doocheol. 1988. "The Determinants of Life Insurance Growth in Developing Coun-
tries, with Particular Reference to the Republic of Korea." Ph.D. dissertation. Geor-
gia State University, College of Business Administration, Atlanta.
La Porta, Rafael, Florencio L�pez-de-Silanes, Andrei Shleifer, and Robert W. Vishny.
1999. "The Quality of Government." Journal of Law, Economics, and Organization
15:222�79.
Lewis, Frank D. 1989. "Dependents and the Demand for Life Insurance." American
Economic Review 79:452�66.
Loayza, Norman, Humberto Lopez, Klaus Schmidt-Hebbel, and Luis Serven. 1999. "World
Saving Database." World Bank, Development Research Group, Washington, D.C.
Lundberg, Mattias, and Lyn Squire. 2001. "Growth and Inequality: Extracting the Les-
sons for Policymakers." World Bank, Development Research Group, Washington, D.C.
Meng, Xingguo. 1994. "Insurance Markets in Developing Countries: Determinants, Policy
Implications, and the Case of China." Ph.D. dissertation. Temple University, Fox
School of Business and Management, Philadelphia.
Outreville, Fran�ois J. 1992. "The Relationship between Insurance, Financial Develop-
ment and Market Structure in Developing Countries." UNCTAD Review 3:53�69.
------. 1996. "Life Insurance Markets in Developing Countries." Journal of Risk and
Insurance 63:575�93.
------. 1999. "Financial Development, Human Capital, and Political Stability." unctad
Discussion Paper 142. United Nations Conference on Trade and Development, Geneva.
88 the world bank economic review, vol. 17, no. 1
Political Risk Services. Various years. International Country Risk Guide. Syracuse, N.Y.
Skipper, Harold, Jr., and Kenneth Black Jr. 2000. Life Insurance, 13th ed. Englewood
Cliffs, N.J.: Prentice Hall.
Soo, Hak Hong. 1996. "Life Insurance and Economic Growth: Theoretical and Empiri-
cal Investigation." Ph.D. dissertation. University of Nebraska, Department of Econom-
ics, Lincoln.
Swiss Reinsurance Company. Various years. Sigma. Zurich.
Truett, Dale B., and Lila J. Truett. 1990. "The Demand for Life Insurance in Mexico and
the United States: A Comparative Study." Journal of Risk and Insurance 57:321�28.
unctad (United Nations Conference on Trade and Development). 1982. The Promo-
tion of Life Insurance in Developing Countries. TD.B.C.3/177. Geneva.
undp (United Nations Development Programme). 2002. Human Development Report
2002. New York: Oxford University Press.
United Nations Statistics Division. 1993. System of National Accounts 1993. New York.
Ward, Damian, and Ralf Zurbruegg. 2000. "Does Insurance Promote Economic Growth?
Evidence from oecd Countries." Journal of Risk and Insurance 67:489�506.
Webb, Ian. 2000. "The Effect of Banking and Insurance on the Growth of Capital." Ph.D.
dissertation. Georgia State University, College of Business Administration, Atlanta.
World Bank. Various years. World Development Indicators. Washington, D.C.
Yaari, Menahem E. 1965. "Uncertain Lifetime, Life Insurance, and the Theory of the
Consumer." Review of Economic Studies 32:137�50.
Zelizer, Vivian R. 1979. Morals and Markets: The Development of Life Insurance in the
United States. New York: Columbia University Press.
the world bank economic review, vol. 17, no. 1 89�106
Benefits on the Margin:
Observations on Marginal Benefit Incidence
Stephen D. Younger
Benefit incidence analysis has become a popular tool over the past decade, especially
for researchers at the World Bank. Despite or perhaps because of the popularity of this
method, recent research has pointed out many of its limitations. One of the most com-
mon criticisms of benefit incidence analysis is that its description of average participa-
tion rates is not necessarily useful in guiding marginal changes in public spending
policies. This article considers a variety of methods for analyzing the marginal benefit
incidence of policy changes. A key conceptual point is that despite the fact that the
various methods measure "marginal" incidence, they do not measure the same thing--
nor are they intended to do so. There are many possible policy changes and thus many
margins of interest. Each method captures one of these and so is of interest for some
analyses and inappropriate for others. Empirically, the precision of the methods dif-
fers substantially, with those relying on differenced data or aggregations of households
yielding standard errors that are quite large relative to the estimated shares.
The past decade has seen a resurgence of interest in the relationship between
poverty and public spending in developing economies. This resurgence has fos-
tered the return of incidence analysis, particularly for the benefits of public spend-
ing in the social sectors. Although analysis of tax incidence has a long and
venerable history in economics, distributional analysis of the benefits of public
spending--and public policy more generally--is more recent (Aaron and McGuire
1970; Brennan 1976; Meerman 1979; Selowsky 1979). Broadly stated, benefit
incidence analysis assesses how the benefits of government spending are distrib-
uted across the population.
Though there are many ways to approach this issue, a fairly standard method
has emerged, largely based on the work of researchers at the World Bank (Demery
1997; van de Walle and Nead 1995; Selden and Wasylenko 1992). This method
takes "across the population" to mean "across the expenditure (or income) dis-
Stephen D. Younger is Associate Director of the Food and Nutrition Policy Program at Cornell
University. His e-mail address is sdy1@cornell.edu. The author is grateful for helpful comments from
Fran�ois Bourguignon, Peter Glick, and participants in the Organisation for Economic Co-operation
and Development research program "Development of Human Resources and Poverty Alleviation." Some
of the research reported in this article was supported by the U.S. Agency for International Development
under cooperative agreement AOT -0546-A-00-3178-00.
DOI: 10.1093/wber/lhg009
� 2003 The International Bank for Reconstruction and Development / THE WORLD BANK
89
90 the world bank economic review, vol. 17, no. 1
tribution"--an approach consistent with the overall concern with poverty. It then
estimates the distribution of benefits based on some variant of the average partici-
pation rate in a public program among people in different expenditure brackets.
Given a presumed preference for public spending that benefits the poor, pro-
grams and policies are viewed more favorably if poor people's average partici-
pation in them is higher than that of the nonpoor. A large increase in the
availability of nationally representative, multipurpose surveys--such as the Liv-
ing Standards Measurement Surveys (Grosh and Glewwe 1998)--and the rela-
tive ease with which this standard method can be applied have led to a profusion
of such analysis. Indeed, benefit incidence analysis has become a common fea-
ture of developing countries' poverty profiles and of many project proposals
and evaluations.
Despite or perhaps because of the popularity of this method, recent research
has pointed out many of its limitations (van de Walle 1998; Lanjouw and
Ravallion 1999). Among the most common criticisms of standard benefit inci-
dence analysis is that its description of average participation rates is not neces-
sarily useful in guiding marginal changes in public spending policies--a point
first made by Lipton and Ravallion (1995). The logic of this argument is com-
pelling. The standard method describes who is currently benefiting from a par-
ticular public expenditure. As such, it is a useful guide to the effects of a policy
change that distributes benefits in proportion to current benefits.
But a policy change that increases spending will not necessarily go to existing
beneficiaries in proportion to their current benefits--or even go to existing ben-
eficiaries at all. Many policies explicitly aim to expand the benefits of public
spending among nonbeneficiaries. In such cases, because the benefits do not go
to existing beneficiaries, the standard method is misleading. Even if services
change for existing beneficiaries, the changes may not be uniform, in which case
the standard method is also inappropriate. For example, a policy to ensure that
all students have a complete set of textbooks will have different distributional
consequences if some students already have a complete set, and so gain nothing,
whereas others do not.
In response to such observations, several recent studies have proposed alter-
native methods to measure the marginal incidence of public spending. Glick and
Razakamanantsoa (2001) and Younger (2002) examine shares of the change in
benefits over time across the expenditure distribution. Lanjouw and Ravallion
(1999) and Galasso and Ravallion (2001) estimate the "marginal odds of par-
ticipation" for each expenditure quantile as the coefficient in a regression of
quantile and small area participation rates on large area participation rates.
Lanjouw and others (2002) and Ravallion (1999) apply similar techniques to
panel data to control for fixed area characteristics. Younger (1999, 2002) con-
siders marginal incidence to be the distribution of compensating variations for
marginal policy changes, based on estimated demands for public services.
This article considers each of these approaches to analyzing the marginal
benefit incidence of policy changes using a specific example of secondary educa-
Younger 91
tion in rural Peru. A key conceptual point is that despite the fact that all of these
methods claim to measure "marginal" incidence, they do not measure the same
thing--nor are they intended to do so. There are many possible policy changes,
and thus many margins of interest. Each method captures one of these and so is
appropriate for some analyses and inappropriate for others. Empirically, the
precision of the methods differs substantially, with those that rely on differenced
data or aggregations of households into groups yielding standard errors that are
quite large relative to the estimated shares. This result argues for caution with
these methods when using samples similar to those in the Peru surveys, which
are about the same size as many existing multipurpose household surveys.
I. Methods
This section presents six alternative methods to measure the marginal incidence
of a public policy change and compares them to the standard benefit incidence
method.
The Standard Benefit Incidence Method
A standard benefit incidence study requires two components: a measure of the
value of the benefit that an individual, household, or population group receives
from a particular public expenditure; and a way to compare the beneficiaries to
the general population. When studying the benefits of public services, the stan-
dard method usually uses the government's cost of provision to estimate a service's
value to users. But there are both theoretical and practical reasons to doubt this
practice (van de Walle 1998; Sahn and Younger 2000). So an increasing num-
ber of evaluations simply count users--that is, a user or beneficiary gets a bene-
fit of one, others get zero.
It is possible to compare the beneficiaries of a public expenditure with the
general population along many dimensions, including ethnicity, gender, region
of residence, age, functional income classifications, or political constituency. But
an interest in poverty and inequality implies that most comparisons will involve
welfare. That is, the goal is to know how the recipients' welfare compares with
the general population's welfare. Almost all work analyzing developing econo-
mies uses household expenditures per capita or per adult equivalent as its mea-
sure of welfare. Once it has been decided how to value benefits and how to group
the sample, the calculations are simple: divide each individual's or household's
benefit by the total to get his or her share of benefits, and sum those shares across
a population group, usually welfare quantiles.
The standard method clearly uses group averages to estimate the distribution
of benefits. Nevertheless, this average measure does yield the distributional con-
sequences of a marginal policy change that distributes benefits to existing users
in proportion to their benefit. An obvious example of such a policy change is a
tax or subsidy that changes an existing price proportionately. But one can think
of others, such as a new uniform for each child in school or a new vaccination
92 the world bank economic review, vol. 17, no. 1
for each member of a social security system. As such, the terminology that com-
pares "average" benefit incidence, calculated in the standard way, to "marginal"
benefit incidence, calculated with one of the other methods mentioned in the
introduction, is unfortunate. The standard method does capture a margin and
can be interpreted as such in terms of welfare theory (Yitzhaki and Slemrod 1991).
Rather, the problem with the standard method is that this is not the margin that
interests most people. For example, policymakers often do not think of increases
in public spending for health or education in terms of larger price subsidies for
those services. Instead they have in mind an expansion of these services to non-
beneficiaries induced by increased access or reduced rationing rather than reduced
price. As noted, by definition such benefits do not go to existing beneficiaries,
so the standard method is inappropriate. The next section discusses several
methods for estimating the distributional effects of nonproportional expansions
of public service coverage.
Estimating the Benefits of a Marginal Expansion in Services
Method 1A: Using Spatial Variations in Coverage to Estimate Mar-
ginal Program Benefits. Lanjouw and Ravallion (1999) develop a political
economy model in which different population groups--such as the poor and
nonpoor--have different political power and different costs and benefits from a
given public expenditure. (Similar models are found in Ravallion 1999, 2002.)
The interplay between these factors determines the relationship between the size
of a program or service, total public spending on it, and each group's share of its
benefits. "Early capture" by the poor occurs when they receive larger shares of
a small program but their share declines as the program grows.1 "Late capture"
is the opposite case.
Even with substantial restrictions, the theoretical model yields no general re-
sults on whether early or late capture will occur, so the question requires em-
pirical analysis. To that end, Lanjouw and Ravallion estimate the following
regression:
(1) pi,k,q = aq + bqpk + uq
where i indexes a small geopolitical unit (a province in Peru), k indexes a larger
one (a department in Peru), and q indexes the welfare quantile. The left-hand
variable is the program participation rate for a given province and quantile. The
regressor is the program participation rate for the department in which that
province is located. bq, then, is the marginal effect of an increase in the program
participation rate for the department on the participation rates of people in a
1. In Lanjouw and Ravallion's specification, the nonpoor bear all the program costs and hold all
the political power in the sense that the poor cannot impose on them a program that lowers their wel-
fare. In such cases the convexity of the program cost function is sufficient to guarantee "early capture"
by the poor.
Younger 93
given province and quantile. The regression is run separately for each quantile.
In addition, because pi,k,q is included in pk, there is an upward bias in the esti-
mate of bq. Lanjouw and Ravallion resolve this by instrumenting pk with the left-
out mean--that is, the participation rate for all of department k except those
individuals in province i and quantile q.
The intuition behind the regression is that by observing variations in depart-
mental participation across the country, it is possible to understand how increased
coverage affects the participation of different population groups. If bq is greater
than one, it indicates that a general expansion in coverage is correlated with a
disproportionately large increase in participation for that province and quantile.
One advantage of this method is that it requires only a cross-section of data, just
like the standard method. An important assumption is that across regions, the
same political process determines the correlation between program size or cov-
erage and incidence.
The margin that this model estimates is the incidence of an increase in pro-
gram participation. The model does not address the policies that might bring
about the program expansion, nor does it consider specific changes in the de-
mand for services. Rather, it makes a more general appeal to the political economy
behind the policies to argue that, whatever policies are used--price reductions,
quality improvements, reduced rationing--the outcome must respect the politi-
cal constraints implied by each group's costs, benefits, and political power.
Method 1B: Controlling for Fixed Effects. Lanjouw and Ravallion
(1999) point out that equation 1 includes no controls for any effects on prov-
ince and quantile participation rates except the department's participation rate.
Where data are available for more than one point in time, it is possible to con-
struct a panel of provinces and thus to include a province fixed effect to control
for left-out covariates that are constant over time. (Lanjouw and others 2002
and Ravallion 1999 pursue this strategy.) This is possible even if the surveys are
not panels of households, as long as the households are sampled from the same
provinces and each survey is representative at the province level.2
Method 1C: Using Disaggregated (Individual) Data. A purely statisti-
cal problem with the Lanjouw-Ravallion model is that it uses average data for
provinces and quantiles. Although this approach was often necessary in the past,
individual-level data sets are now available to estimate benefit incidence. Group-
ing observations into province and quantile averages reduces the efficiency of
the estimates, yielding larger estimated standard errors (Johnston 1972). The
2. The first condition is often true of household surveys, whereas the second is quite rare. In the two
Peru surveys used in the next section, there is considerable overlap between provinces, but the samples
are not representative at the province level--leading to the possibility that the observed variation over
time is due to sampling differences. In Peru, however, survey teams usually return to the same clusters
when they conducts new surveys, which should minimize this problem.
94 the world bank economic review, vol. 17, no. 1
application below estimates the model for both group average and individual
data.
Method 2: Observing Changes as Programs Expand over Time. This
method addresses the same margin as method 1, the incidence of increased spend-
ing as a program or service expands. But rather than using the spatial variation
in the correlation between program size and incidence, this method calculates
each group's share of observed changes in benefits. As such it is mechanically
similar to the standard benefit incidence method--except that it substitutes the
change in a given quantile's benefits (or program participation) for its level. Glick
and Razakamanantsoa (2001) and Younger (2002) use two cross-sections at dif-
ferent points in time to estimate each quantile's share in the change in use of
various public services. Van de Walle 1995, Hammer and others 1995, and
Lanjouw and others 2002 also use two cross sections, but they describe how the
standard benefit incidence changes over time rather than the incidence of the
changes.
This method requires at least two cross-sectional surveys. But just as the num-
ber of developing economies with at least one nationally representative multi-
purpose survey grew over the past decade, an increasing number of countries
now have more than one such survey a few years apart. Like method 1, this
approach says nothing about the incidence of program expansions brought about
by particular policy instruments. It is purely a description of what actually took
place between the two surveys in terms of program coverage and shares.
Method 3A: Econometric Estimates of Compensating Variations.
Rather than use the standard benefit incidence method as an approximation for
the compensating variation for a price change, it is possible to estimate, econo-
metrically, compensating variations for price and other policy changes. A well-
established literature in transport and environmental economics does this for
goods and services where demand is discrete (Small and Rosen 1981; McFadden
1995), and Paul Gertler and his associates have applied these techniques to the
demand for health and education in developing countries (Gertler and others
1987; Gertler and Glewwe 1990).
The model is well known. Using education as an example, assume that each
household has a utility function that depends on its consumption and its choice
for the quality of schooling:
(2) Vj = f[y � pj,Q(Xj,Z)] + ej
where j indexes the choice (school or no school), y is household permanent in-
come proxied by household expenditures, pj is the price of choice j including all
opportunity costs of time, and Q is a function that measures quality, which de-
pends on choice-specific characteristics Xj and household or personal character-
istics Z. Households choose the option j that yields the highest utility. Although
Vj is not observable, if a household chooses option j, Vj is greater than all other
Younger 95
Vi. The model estimates the probability that this is the case using only the ob-
served choice, and takes the probability of choosing option j as an expected
demand for that option. Small and Rosen (1981) show how to calculate com-
pensating variations in such a model.
An important identifying assumption of a model like equation 2 is that the
observed choice is the one that provides the highest utility, which implies that
there is no rationing beyond what can be captured with the choice-specific char-
acteristics Xj. But that may not be the case for some public services. For example,
schools may exclude students based on merit, gender, social status or connec-
tions, and so on. In such cases some school-age children may not be attending
school even though that option would provide them with the greatest utility,
which biases the estimates in equation 2 if the rationing is correlated with the
regressors.3 In the case study that follows, rationing is not a problem. Except for
a few prestigious schools in urban areas, secondary schools in Peru do not ra-
tion slots. This is possible because class sizes are not limited.
Unlike methods 1 and 2, this method permits traditional policy analysis in
the sense that it answers the question, "Who will receive the marginal benefits if
policy variable xj is changed?" As such it is useful for estimating the marginal
incidence of any policy change for which appropriate xj data are available.
Method 3B: Econometric Estimates of Changes in the Probability
of Participation. The method based on compensating variations differs from
the other methods presented in that it considers the value of a policy change to
potential recipients rather than the change in the probability of participation.
This valuation adds an extra dimension not found in the other methods. Glick
and Sahn (2000) estimate the model in equation 2 but calculate only the change
in the probability of participation associated with simulated policy changes. By
modeling participation rather than its monetary value, this approach is closer to
the others presented than method 3A. An advantage of this method over the es-
timation of compensating variations is that because it models only the probabil-
ity of a given option, it remains valid in the presence of rationing.
II. Benefit Incidence of Secondary Schooling in Rural Peru
This section estimates the distribution of benefits from an expansion of secondary
schooling in rural Peru using all the methods outlined in the previous section. In
addition, it calculates standard measure of benefit incidence. The calculations
for shares in observed changes in participation--that is, method 2--are straight-
forward. Note that secondary schooling coverage rates increased slightly in rural
3. Some nonprice rationing that is characteristic of the service can, however, be modeled. For ex-
ample, health centers may charge low fees and handle the excess demand by imposing long waiting
times. As long as the waiting time can be included in Xj, the estimates are consistent.
96 the world bank economic review, vol. 17, no. 1
Peru between 1994 and 1997, from 6.33 to 6.58 percent of the rural popula-
tion. The other methods require preliminary regressions, which are reported first.
Estimating the Marginal Odds of Participation
The results of four regressions based on the Lanjouw and Ravallion (1999) model,
as described in method 1, are reported in table 1 and table 2. The cross-sectional
data are for rural households from the 1994 and 1997 rounds of the Encuesta
Nacional de Hogares sobre Medici�n de Niveles de Vida (the National House-
hold Living Standards Measurement Survey). The regressions are for rural prov-
inces only, to be consistent with the model for method 3. In the Individual-level
model section of table 1 the dependent variable is a zero/one indicator of
individual-level participation in secondary schooling. (Thus, the linear probability
model is being estimated; it would also be possible to estimate this model as a
probit or logit.) In the other sections of tables 1 and 2 the data for the dependent
variable are the province- and quintile-specific participation rates for secondary
school attendance, defined as the number of secondary students divided by the
population.
In all the models the right-hand variable is the department-wide participation
rate, and all estimates are two-stage least squares using the left-out mean de-
partment participation rate as an instrument. In addition, two restrictions have
been imposed on the coefficients: that the aqs sum to zero and the bqs sum to the
number of quantiles, five in this case. Although Lanjouw and Ravallion do not
impose these restrictions, they are required if the estimated shares of marginal
benefits are to sum to one. As it turns out the unrestricted estimates are quite
close to those reported here for all the cross-sectional models. For the panel data
model the differences are much larger--but so are the standard errors, so even
Table 1. Estimates for the Lanjouw-Ravallion (1999) Model Using Province-
and Individual-Level Cross-Sectional Data, 1994
Province-level model, 1994 Individual-level model, 1994
Quintilea Coefficient SE t-statisticb N Coefficient SE t-statisticb N
1 Intercept �1.293 1.129 �1.145 9.785 2.249 4.351
Slope 0.688 0.136 �2.302 61 0.387 0.059 �10.469 210
2 Intercept 0.652 1.400 0.466 3.719 2.893 1.285
Slope 0.778 0.179 �1.246 68 0.912 0.079 �1.106 227
3 Intercept �0.219 1.703 �0.128 �9.801 2.861 �3.426
Slope 1.241 0.216 1.119 62 1.493 0.084 5.895 258
4 Intercept �0.315 1.606 �0.196 3.136 2.939 1.067
Slope 1.210 0.212 0.989 72 1.003 0.089 0.030 280
5 Intercept 1.175 1.652 0.711 �6.839 2.825 �2.421
Slope 1.084 0.205 0.407 67 1.205 0.091 2.267 363
aQuintiles are based on household expenditures per capita.
bThe t-statistics test against zero for the intercept and one for the slope.
Source: Author's calculations
Younger 97
Table 2. Estimates for the Lanjouw-Ravallion (1999) Model Using Province-
Level Cross-Sectional and Panel Data, 1994�97
Province-level model, 1997 Province�level model, 1994-97 panela
Quintileb Coefficient SE t-statisticc N Coefficient SE t-statisticc N
1 Intercept 2.950 1.689 1.747
Slope 0.250 0.218 �3.443 62 0.611 0.217 �1.797 38
2 Intercept 1.067 1.612 0.662
Slope 0.888 0.211 �0.532 78 0.595 0.417 �0.972 49
3 Intercept �0.179 1.892 �0.095
Slope 1.039 0.249 0.158 76 0.908 0.711 �0.130 44
4 Intercept �2.019 1.751 �1.153
Slope 1.454 0.230 1.976 76 0.579 0.479 �0.880 50
5 Intercept �1.820 1.799 �1.011
Slope 1.369 0.242 1.528 77 2.307 1.735 0.753 48
aIntercept coefficients are province-specific in the panel and thus suppressed. Estimates are for con-
strained models, with intercepts summing to zero and slopes to five.
bQuintiles are based on household expenditures per capita.
cThe t-statistics test against zero for the intercept and one for the slope.
Source: Author's calculations
for this model the slope coefficients in the restricted and unrestricted models do
not differ at the 5 percent level.
Three points about table 1 and table 2 are worth noting here (see later discus-
sion of the distributional implications). First, the various models produce quite
different estimates for the quintile-specific marginal odds of participation. This
is true even for the two cross-sectional models (Province-level model, 1994 and
1997). These differences cannot be due to changing marginal odds of participa-
tion as coverage expands, because in these samples rural secondary school enroll-
ments were essentially constant between 1994 and 1997. Although the standard
errors are large, there are several significant differences--especially for the poorest
quintile's marginal share. Second, in all the models the poorer quintiles receive
a less than proportionate share of marginal benefits from secondary schooling.
Finally, as expected, the standard errors for the slope coefficients in the indi-
vidual-level model are only about half those of the province-level model.
Estimating Demand
Estimates of the demand for secondary schooling in rural Peru, using the same
1994 data set, are shown in table 3. (The 1994 data are used because that sur-
vey asked a broader range of questions about school quality, including distance
to school and parents' evaluations of problems at their children's schools.)
Though it is customary to consider multiple options for schooling--no school,
public school, or private school, or no school, local school, or distant school--
in rural Peru only 3 percent of children attend private school, making an esti-
mate of the demand for private options infeasible. Moreover, the survey does
not identify students who are at school away from home. Thus, the table esti-
98 the world bank economic review, vol. 17, no. 1
Table 3. Probit Estimates for Secondary School Choice in Rural Peru, 1994
Regression results Data
Variable Coefficient SE t-statistic Mean SE
Household characteristics
Constant �1.4144 0.9083 �1.56 1.00 0.00
Net expenditures/1000a 0.7007 0.2606 2.69 7.81 5.58
Net expenditures/1000 squareda �0.0037 0.0030 �1.23 92.16 191.82
Net expenditures � distance 0.0016 0.0017 0.98 34.38 53.31
Age of household head/10 0.1029 0.0463 2.22 4.70 1.18
Gender of household head 0.0250 0.1400 0.18 0.09 0.29
Household head born in urban area 0.2934 0.1019 2.88 0.31 0.46
Household head years of schooling/10 1.6635 0.3223 5.16 0.43 0.35
Household head years �0.4724 0.2249 �2.10 0.31 0.45
of schooling/10 squared
Household members ages 0�5 �0.1262 0.0456 �2.77 0.92 1.03
Household members ages 6�12 �0.1691 0.0485 �3.49 0.95 0.94
Household members ages 13�18 0.1259 0.0491 2.57 1.58 0.97
Household members ages 19�60 0.0704 0.0376 1.87 2.45 1.20
Household members over 60 �0.0465 0.0948 �0.49 0.22 0.51
Child characteristics
Age/10 �0.0909 0.9591 �0.09 1.58 0.31
Age/10 squared 0.3891 0.2772 1.40 2.58 1.10
Gender �0.1558 0.0948 �1.64 0.50 0.50
Indigenous �0.3083 0.1630 �1.89 0.32 0.47
Indigenous � gender �0.0061 0.1546 �0.04 0.16 0.37
Married �0.3689 0.3453 �1.07 0.09 0.28
Married � gender �1.9072 0.5291 �3.60 0.07 0.25
Child of household head 0.4977 0.4975 1.00 0.83 0.38
Spouse of household head 1.1789 0.6529 1.81 0.04 0.20
Other household member 0.4056 0.5061 0.80 0.12 0.32
School characteristics
Number of required textbooks 0.1287 0.0472 2.73 1.71 0.91
Distance to school �0.0945 0.0266 �3.56 4.41 7.03
Distance to school squared 0.0013 0.0005 2.53 68.85 1312.05
Time to school �0.0020 0.0017 �1.18 37.64 36.69
Primary repetition rate �0.4951 0.2305 �2.15 0.36 0.20
Secondary repetition rate 1.4337 0.6125 2.34 0.03 0.06
Percentage of parents expressing a desire to improve
School building �0.0445 0.2126 �0.21 0.66 0.21
Desks and services �0.3150 0.1324 �2.38 0.48 0.32
Feeding programs �0.1293 0.2100 �0.62 0.23 0.21
Class size �0.2476 0.3831 �0.65 0.05 0.09
Teacher training �0.2070 0.1708 �1.21 0.38 0.26
Teaching materials �0.2743 0.2549 �1.08 0.17 0.15
Library �0.2363 0.1856 �1.27 0.32 0.23
Director's power 0.1301 0.6849 0.19 0.02 0.06
Auxiliary personnel training �0.3209 0.3508 �0.91 0.04 0.08
Other 0.0759 0.2022 0.38 0.26 0.24
aThese coefficients are constrained to be equal across the schooling or no schooling options. All
others are for the differential utility of choosing the schooling option.
Source: Author's calculations
Younger 99
mates only the choice between attending or not attending school, as a probit.
The analysis is limited to rural areas because a model with only a few choices is
inappropriate for most urban areas. A resident of Lima (Peru's capital) has a
choice of many schools, public or private. No survey in Peru makes it possible
to adequately identify (let alone model) these choices.
Because the probit can identify the model only up to the differences in Vj, the
model must normalize against one option, which here will be the no school choice.
Thus, it is assumed that Q(X0,Z) = 0. In the estimates it is assumed that the func-
tion f() in equation 2 is quadratic in net expenditures and constrained to be the
same for each option. There is some debate about the second restriction in the
literature (Dow 1999). But as Gertler and Glewwe (1990) note, it is necessary to
get a sensible estimate of the marginal utility of income, which in turn is neces-
sary to calculate compensating variations (Small and Rosen 1981). The func-
tion Q() is linear and separable from net expenditures except for an interaction
between net expenditures and distance from school. This variable is the one that
will be used to compare results with the Lanjouw-Ravallion method, so the aim
is to allow as much flexibility as possible.4
The sample includes all rural children attending secondary school or eligible
to attend. The latter group includes all children of the appropriate age who
have not graduated from secondary school--including children who have not
graduated from primary school, because in the context of long-run optimization
the decision not to complete primary school is partly affected by perceptions of
the value of secondary school. Dow (1999) defends this type of unconditional
estimate.
All the household characteristics in table 3 are self-explanatory except for net
household expenditures. Household expenditures are defined in the broadest way
possible, including imputed values of owner-occupied housing and own-produced
goods (Younger 2002). In addition, if a household contains a secondary student,
the costs of schooling--including fees, books, uniforms, and transportation--
and the opportunity costs of time at school are added to get a gross expenditure
variable that is before the costs of schooling. The price is the cluster- or district-
level mean for these school costs. (Districts are the third-level geopolitical unit
in Peru, smaller than provinces.) If a cluster has at least four observations, clus-
ter-level data are used to calculate means. Otherwise district-level data are used.
(The same criterion of at least four observations is used for all subsequent clus-
ter- or district-level regressors.) All means calculated in this way are left-out
means. Net expenditures are gross expenditures minus this price variable.
The child characteristics in table 3 are also self-explanatory. The default op-
tion for relation to the household head is being the head. The school character-
istics require some explanation. The number of required books is the cluster- or
4. As it turns out, the estimated changes in probability of attendance and compensating variations
from this model have correlation coefficients greater than 0.95 with those from a model without the
interaction.
100 the world bank economic review, vol. 17, no. 1
district-level mean number of textbooks that school requires. Here this is taken
as an indicator of academic quality. Distance to the school is measured in kilo-
meters and time in minutes.
The questions about parents' wish to change features of their school are based
on the following question to one adult per household, if the household includes
children: "If you could change anything about your children's school, what
changes would you make? (Use a scale from 1 to 3)." This is followed by a list
of school characteristics. The regressors in table 3 are the cluster- or district-
level share of parents expressing in interest in improving each feature.
The signs of the coefficients are almost all consistent with prior expectations.
Net household expenditures have a positive and only slightly concave effect on
the probability of secondary school attendance in rural areas. Children are more
likely to attend secondary school if they live in a household whose head is older,
urban-born, and more educated. Children in households with younger children
(age 12 and under) are less likely to attend secondary school, whereas those in
households with older children (excluding the child being observed) or adults
are more likely to attend.
Of the child characteristics, only one variable is statistically significant at stan-
dard levels: married girls are less likely to attend secondary school even after
accounting for the positive effect resulting from being the household head's spouse
(almost all of whom are women). Being female and coming from a household
with at least one indigenous-language speaker also lowers the probability of
attendance, but the t-statistics are smaller for these variables. All but one of the
school characteristics has significant effects on the probability of attendance. On
the other hand, only one of the parental opinion questions is significantly differ-
ent from zero--that expressing a desire for better desks and services.
The results in table 3 are used to simulate two policy changes. The first is a
reduction in school fees of 100 soles--about the sample average expenditure per
student on school fees, books, uniforms, and transportation. This is a policy
change that the standard benefit incidence method approximates, so the distri-
bution of estimated benefits from the two methods should be close.
The second policy simulation reduces each student's distance to a secondary
school to a maximum of 2 kilometers. This affects about two-thirds of the sample.
School placement is clearly a policy variable and one that many people have in
mind when they think of a policy to expand access to public schools. As such it
is the variable in the demand function most consistent with methods 1 and 2.
For each simulation calculations are made for each child of the compensating
variations for the policy change (method 3A) and of the change in the probabil-
ity that he or she attends school (method 3B).
McFadden (1995) shows that the standard method for calculating compensat-
ing variations in the discrete choice model developed by Small and Rosen (1981)
will yield biased estimates if utility is a nonlinear function of income, as it is in this
model. So, here the compensating variations are calculated using the simulation
method described by McFadden with 1000 repetitions for each observation.
Younger 101
Results
The estimated marginal benefits from secondary schooling to each per capita
expenditure quintile in rural Peru, calculated using each of the methods pre-
sented--along with method 4, which is a standard benefit incidence analysis--
are shown in table 4.5 The shares for the Lanjouw and Ravallion (1999) methods
are those presented in tables 1 and 2 divided by five.
Consider first the estimated distributions of marginal benefits associated with
a program expansion--that is, methods 1A, 1B, 1C, and 2 in columns A through
D. As tables 1 and 2 showed, the results for the Lanjouw-Ravallion methods
(1A, 1B, 1C) are statistically different from equal shares, with a modest antipoor
bias. (All tests are at the 5 percent confidence level.) Only a few of the estimates
differ by economically important amounts, but except for the difference in the
first quintile for methods 1A and 1C, these differences are not statistically sig-
nificant due to the relatively large standard errors. This is especially the case
for the panel data method (1B), which estimates a much larger marginal share
for the fifth quintile than the other methods--but with a standard error so large
that the estimate is not distinguishable from either zero or one.
The results for method 2 in column D show an extremely progressive distri-
bution of marginal shares, with the first two quintiles capturing more than 100
percent of the change in benefits. This result is possible because a quintile can
have negative marginal benefits--that is, a decline in its participation rate over
time--even as overall participation increases. (This was true for the third and
fifth quintiles between 1994 and 1997 in rural Peru.) Method 2 seems unsatis-
factory on two counts. First, the marginal share estimates are very different from
all of the method 1 estimates and also quite erratic. The latter phenomenon can
be explained by the fact that the denominator of the marginal shares--the change
in secondary school attendance between 1994 and 1997--is very small: only 0.25
percent of the rural population. With such a small overall change, any quantile's
share of that change can be large even if its participation did not change much.
That said, method 2 is meant to capture shares of marginal changes, which by
definition are small. Thus, applying this method to services with larger expan-
sions might produce more stable estimates, but such estimates would be less
accurately termed marginal.
A more important problem with method 2 is that, because it relies on differenced
data, the estimates have very large standard errors--so large that the marginal share
estimates for this method are statistically indistinguishable from the others despite
its very different point estimates. Though this is only one example, it seems that
5. The quintiles are based only on the rural samples of the Peru surveys. It would be just as easy to
derive them for all households in the sample and give zero benefits to urban residents. Where rural
residents fall in the nationwide spending, distribution would then influence the estimated shares--in
particular, subsidies to rural secondary schools would look more progressive because rural residents
are poorer than urban residents in Peru--but comparisons across methods of each quantile's share would
not change.
Table 4. Quintile Shares of Marginal Benefits to Secondary Schooling in Rural Peru, Calculated Using Various Methods
Method 1A Method 1B Method 1C Method 2 Method 3Aa Method 3Ab Method 3Ba Method 3Bb Method 4
Quintile A B C D E F G H I
1 0.14 0.12 0.08 0.36 0.16 0.13 0.21 0.15 0.10
(0.027) (0.043) (0.012) (1.035) (0.048) (0.010) (0.023) (0.010) (0.014)
2 0.16 0.12 0.18 0.81 0.16 0.18 0.22 0.20 0.17
(0.036) (0.083) (0.016) (1.824) (0.047) (0.011) (0.026) (0.011) (0.016)
3 0.25 0.18 0.30 �0.19 0.20 0.21 0.22 0.22 0.23
(0.043) (0.142) (0.017) (1.425) (0.051) (0.012) (0.026) (0.012) (0.019)
102
4 0.24 0.12 0.20 0.44 0.20 0.24 0.20 0.24 0.24
(0.042) (0.096) (0.018) (1.149) (0.056) (0.013) (0.027) (0.012) (0.020)
5 0.22 0.46 0.24 �0.43 0.29 0.24 0.15 0.20 0.26
(0.041) (0.347) (0.018) (1.880) (0.055) (0.015) (0.022) (0.012) (0.022)
Note: Numbers in parentheses are standard errors. The methods are 1A: Lanjouw and Ravallion (1999) model applied to 1994 cross-section of province-
level data. 1B: same as 1A but applied to a panel of provinces, in 1994 and 1997, with fixed effects. 1C: same as 1A but applied to individual-level data. 2:
shares of observed changes in secondary school attendance, 1994�97. 3A: shares of estimated compensating variations associated with a policy change, indi-
vidual-level data, 1994. 3B: shares of estimated change in probability of attendance associated with a policy change, individual-level data, 1994. 4: shares
based on standard benefit incidence, using a zero or one indicator of attendance. Quintiles are based on household expenditures per capita.
aSimulation of a reduction in distance to school to a maximum of 2 km.
bSimulation of a reduction in school fees of 100 soles.
Source: Author's calculations.
Younger 103
method 2 may require samples that are much larger than are typical in developing
economies to produce precise estimates of marginal shares.
As noted, increased access is the policy that most people have in mind when
considering a program expansion. For secondary schools in rural Peru, this is
best captured by reduced distance to school--a relevant variable in Peru during
the period being analyzed. In the 1990s Peru's government invested consider-
able resources in building and rehabilitating schools through the Fondo Nacional
de Compensaci�n y Desarrollo Social (the National Compensation and Social
Development Fund) (Paxson and Schady 1999), and in rural areas the median
travel time to get to school fell from 40 to 30 min between 1994 and 1997. (The
1997 survey did not ask for distance to school, only travel time.) So even though
methods 1 and 2 apply to all changes that affect program size, reduction in dis-
tance to school should have been an important factor during this period. As such,
it is interesting to compare method 1 to method 3.
As noted, the policy change simulated here is a reduction in distance to school
to a maximum of 2 km.6 The estimated distributions of the changes in the prob-
ability of attendance (column G, representing method 3B, in table 4) are some-
what more progressive than those for the compensating variations (column E,
representing method 3A), though only the fifth quintile's shares differ signifi-
cantly. These findings indicate that in rural Peru, the value that people place on
secondary schooling increases with household expenditures per capita. Both
estimates of marginal shares are somewhat larger than any of the Lanjouw-
Ravallion methods for the poorest quintile.
For the changes in the probability of attendance (column G) all of these differ-
ences are statistically significant. But for the compensating variations (column E)
only the first quintile difference with method 1C--the most precise of the
Lanjouw-Ravallion methods--is statistically significant. The changes in prob-
ability (column G) also differ significantly from method 1C (the individual-model)
in the third and fifth quintiles, and from method 1A (the province-level cross-
section) in the fifth quintile. For the compensating variations (column E) the only
other significant difference is for the shares for the third quintile relative to method
1C. That the estimates differ for some quintiles is evidence that more is at work
than the distance to school across space and across samples in Peru, which is
not surprising.
Finally, quintile shares estimated with the standard benefit incidence model
(column I in table 4) based on school attendance are quite close to those derived
from both the change in probability of attendance (column H, representing
method 3B) and the compensating variations (column F, representing method
3A) for a 100 sole price change--though the differences at the first quintile are
statistically significant, if minor (0.05 and 0.03, respectively), as is the differ-
6. The mean distance in the sample is 4.4 km and the median is 2.7 km, with a standard deviation
of 7.0 km. The extreme distance is 36.0 km.
104 the world bank economic review, vol. 17, no. 1
ence at the fifth quintile for model 3B (column H). Thus, the standard method
yields a good approximation to the marginal incidence of a price change.
It is also interesting to note that unlike the examples cited in Ravallion (2002)--
including the original Lanjouw and Ravallion (1999) result--the shares from
the standard benefit incidence method for rural secondary school attendance in
Peru are not significantly less propoor than those from any of the other meth-
ods, except for the simulation of reduced distance to school.
III. Conclusion
Benefit incidence analysis is now quite common, partly because of the impor-
tance of the issue that it addresses and partly because it is easy to do. Neverthe-
less, critics have pointed out that the standard method used to carry out the
analysis can often be misleading because it uses quantile average shares of ben-
efits, whereas analysis of any policy change should be done on the margin. This
article argues that the standard method can in fact be interpreted as a marginal
method: it gives a first-order approximation of the distributional consequences
of a price change or any other change that affects only observed beneficiaries in
proportion to their existing benefit. In the example of secondary school atten-
dance in rural Peru, the approximation is reasonably good, even for a large
(nonmarginal) change in the cost of attendance. This finding is consistent with
previous work on five social services in Ecuador (Younger 1999). In that sense
this article supports the standard method--as long as it is interpreted correctly.
More broadly, however, there are more margins of interest than price. In
particular, expanded access to services, rather than changes in fees, is often what
policymakers have in mind when considering increased spending on a public
service. This article explores different methods that apply to different margins.
Methods 1 and 2 do not identify specific causes of a program expansion but rather
argue that however a program expands, it will have to respect political economy
constraints, which can be captured in the correlation between a program's size
and its distribution of benefits. Method 3, on the other hand, is grounded in more
traditional policy analysis, identifying the marginal incidence of a specific policy
change based on a household's compensating variations or willingness to pay for
that change--or, more narrowly, on the change in its probability of participation.
In general, the different methods produce different estimates of marginal bene-
fit incidence, suggesting that analysts should tailor their choice of method to the
issue at hand. Those interested in a general description of program beneficiaries
or in the incidence of change in benefits proportional to existing use can use the
standard benefit incidence method (method 4) or the methods based on de-
mand estimates (method 3). Method 2 is appropriate for those interested in a
description of the incidence of changes in benefits over time. For the distribu-
tional consequences of a general expansion in program coverage in an unchang-
ing political-economic environment, one of the Lanjouw-Ravallion methods
(method 1) is the relevant option. Those interested in analyzing specific policy
Younger 105
changes with effects that are not proportional to existing demand should choose
the methods based on demand analysis (method 3).
Apart from these conceptual differences in methods, two important results
from the examples relate to the precision of each method's estimates. First,
methods that rely on individual- or household-level data yield smaller stan-
dard errors than those that use regional aggregations. Thus, a straightforward
modification of the Lanjouw-Ravallion method using individual-level data is
preferable where such data are available. Second, methods that rely on differ-
ences over time have large standard errors. Though not surprising, this makes it
difficult to use method 2 with existing surveys, few of which have enough obser-
vations to provide adequate precision.
References
Aaron, Henry, and Martin C. McGuire. 1970. "Public Goods and Income Distribution."
Econometrica 38(6):907�20.
Brennan, Geoffrey. 1976. "The Distributional Implications of Public Goods." Econometrica
44(2):391�99.
Demery, Lionel. 1997. "Benefit Incidence Analysis." World Bank, Poverty Reduction and
Economic Management Network, Poverty Anchor, Washington, D.C.
Dow, William. 1999. "Flexible Discrete Choice Demand Models Consistent with Utility
Maximization: An Application to Health Care Demand." American Journal of Agri-
cultural Economics 81(3):680�85.
Galasso, Emanuela, and Martin Ravallion. 2001. "Decentralized Targeting of an Anti-
Poverty Program." World Bank, Development Research Group, Washington, D.C.
Gertler, Paul, and Paul Glewwe. 1990. "The Willingness to Pay for Education in
Developing Countries: Evidence from Rural Peru." Journal of Public Economics
42(3):251�75.
Gertler, Paul, Lu�s Locay, and Warren Sanderson. 1987. "Are User Fees Regressive? The
Welfare Implications of Health Care Financing Proposals in Peru." Journal of Econo-
metrics 33(1/2):67�88.
Glick, Peter, and Mamisoa Razakamanantsoa. 2001. "The Distribution of Social Services
in Madagascar, 1993�99." cfnpp Working Paper 128. Cornell University, Cornell Food
and Nutrition Policy Program, Ithaca, N.Y.
Glick, Peter, and David Sahn. 2000. "The Demand for Primary Schooling in Rural Mada-
gascar: Price, Quality, and the Choice Between Public and Private Providers." cfnpp
Working Paper 113. Cornell University, Cornell Food and Nutrition Policy Program,
Ithaca, N.Y.
Grosh, Margaret E., and Paul Glewwe. 1998. "Data Watch: The World Bank's Living
Standards Measurement Study Household Surveys." Journal of Economic Perspec-
tives 12(1):187�96.
Hammer, Jeffrey, Ijaz Nabi, and James Cercone. 1995. "Distributional Effects of Social
Sector Expenditures in Malaysia, 1974 to 1989." In Dominique Van de Walle and
Kimberley Nead, eds., Public Spending and the Poor. Baltimore, Md.: Johns Hopkins
University Press.
Johnston, J. 1972. Econometric Methods. New York: McGraw-Hill.
106 the world bank economic review, vol. 17, no. 1
Lanjouw, Peter, and Martin Ravallion. 1999. "Benefit Incidence, Public Spending Reforms,
and the Timing of Program Capture." World Bank Economic Review 13(2):257�73.
Lanjouw, Peter, Menno Pradhan, Fadia Saadah, Haneen Sayed, and Robert Sparrow.
2002. "Poverty, Education, and Health in Indonesia: Who Benefits from Public Spend-
ing?" In Christian Morrisson, ed., Education and Health Expenditure and Develop-
ment: The Cases of Indonesia and Peru. Paris: Organisation for Economic Co-operation
and Development.
Lipton, Michael, and Martin Ravallion. 1995. "Poverty and Policy." In Jere Behrman
and T. N. Srinivasan, eds., The Handbook of Development Economics, vol. 3B.
Amsterdam: North-Holland.
McFadden, Daniel. 1995. "Computing Willingness-to-Pay in Random Utility Models."
University of California at Berkeley, Department of Economics.
Meerman, Jacob. 1979. Public Expenditure in Malaysia: Who Benefits and Why? New
York: Oxford University Press.
Paxson, Christina, and Norbert Schady. 1999. "Do School Facilities Matter? The Case
of the Peruvian Social Fund (foncodes)." Policy Research Working Paper 2229.
World Bank, Washington, D.C.
Ravallion, Martin. 1999. "Is More Targeting Consistent with Less Spending?" Interna-
tional Tax and Public Finance 6(3):411�19.
------. 2002. "Who Is Protected? On the Incidence of Fiscal Adjustment." Paper pre-
sented at the International Monetary Fund Conference on Macroeconomic Policies
and Poverty Reduction, Washington, D.C. March.
Sahn, David, and Stephen Younger. 2000. "Expenditure Incidence in Africa: Micro-
economic Evidence." Fiscal Studies 21(3):329�48.
Selden, Thomas M., and Michael J. Wasylenko. 1992. "Benefit Incidence Analysis in
Developing Countries." Policy Research Working Paper 1015. World Bank, Wash-
ington, D.C.
Selowsky, Marcelo. 1979. Who Benefits from Government Expenditures? A Case Study
of Colombia. New York: Oxford University Press.
Small, Kenneth, and Harvey Rosen. 1981. "Applied Welfare Economics with Discrete
Choice Models." Econometrica 49(1):105�30.
Van de Walle, Dominique. 1995. "The Distribution of Subsidies through Public Health
Services in Indonesia, 1978�87." In Dominique Van de Walle and Kimberley Nead,
eds., Public Spending and the Poor. Baltimore, Md.: Johns Hopkins University Press.
------. 1998. "Assessing the Welfare Impacts of Public Spending." World Development
26(3):365�79.
Van de Walle, Dominique, and Kimberly Nead. 1995. Public Spending and the Poor.
Baltimore, Md.: Johns Hopkins University Press.
Yitzhaki, Shlomo, and Joel Slemrod. 1991. "Welfare Dominance: An Application to
Commodity Taxation." American Economic Review 81(3):480�96.
Younger, Stephen D. 1999. "The Relative Progressivity of Social Services in Ecuador."
Public Finance Review 27(3):310�52.
------. 2002. "Public Social Sector Expenditures and Poverty in Peru." In Christian
Morrisson, ed., Education and Health Expenditure and Development: The Cases of
Indonesia and Peru. Paris: Organisation for Economic Co-operation and Development.
the world bank economic review, vol. 17, no. 1 107�131
Reducing Child Malnutrition:
How Far Does Income Growth Take Us?
Lawrence Haddad, Harold Alderman, Simon Appleton,
Lina Song, and Yisehac Yohannes
How rapidly will child malnutrition respond to income growth? This article explores
that question using household survey data from 12 countries as well as data on malnu-
trition rates in a cross-section of countries since the 1970s. Both forms of analysis yield
similar results. Increases in income at the household and national levels imply similar
rates of reduction in malnutrition. Using these estimates and better than historical in-
come growth rates, the article finds that the Millennium Development Goal of halving
the prevalence of underweight children by 2015 is unlikely to be met through income
growth alone. What is needed to accelerate reductions in malnutrition is a balanced
strategy of income growth and investment in more direct interventions.
Great strides have been made in reducing child malnutrition in the past few de-
cades. The prevalence of underweight children under age five in the developing
economies was 37.4 percent in 1980. By 2000 this had dropped to 26.7 percent
(acc/scn 2000). Nevertheless, 150 million children in developing areas remain
underweight, and 182 million remain stunted (low height for age). Moreover,
progress in reducing prevalence rates has slowed in the past two decades, and in
Africa both the number and the prevalence of underweight children have increased.
At current trends it is clear that the goal of halving the prevalence of underweight
children between 1990 and 2015--one of the indicator targets for the Millennium
Development Goals for poverty and hunger--will not be met (acc/scn 2000).
What is needed to accelerate reductions in malnutrition to meet this target?1
It is well accepted that a reduction in income poverty will lead to a reduction in
Lawrence Haddad (l.haddad@cgiar.org) is the Director of Food Consumption and Nutrition Di-
vision, International Food Policy Research Institute. Yisehac Yohannes ( y.yohannes@cgiar.org) is
a Research Analyst with the International Food Policy Research Institute. Harold Alderman
(halderman@worldbank.org) is Lead Human Development Economist in the Africa region of the World
Bank. Simon Appleton (simon.appleton@nottingham.ac.uk) is a Lecturer in Economics, University
of Nottingham, England. Lina Song (l.song@nottingham.ac.uk) is at the Institute of Contemporary
Chinese Studies, University of Nottingham.
1. We note Maxwell's (1999, p. 93) reminder that "international targets can oversimplify and
overgeneralize complex problems . . . and distort public expenditure priorities." But even if one ques-
tions the analytical basis of such targets, the general question of how to hasten improvements in nutri-
tion remains a concern.
DOI: 10.1093/wber/lhg012
� 2003 The International Bank for Reconstruction and Development / THE WORLD BANK
107
108 the world bank economic review, vol. 17, no. 1
malnutrition (Strauss and Thomas 1998). Greater incomes at the household level
mean that families can invest more in food consumption, access to clean water
and good hygiene, and effective health care. They can also afford more effective
child care arrangements. At the community level greater income will eventually
lead to better access to and better quality of health care centers and water and
sanitation systems. But will moderate income growth alone be enough to meet
development targets? If the relationship between income growth and malnutri-
tion reduction is not sufficiently strong, more direct investments will be needed
to accelerate declines in malnutrition. Candidates for such investment include
nutrition programs such as community-based behavior change initiatives and
micronutrient supplementation and fortification (Allen and Gillespie 2001).
The less than perfect correlation between nutritional status and national in-
come levels or national income distribution is often used to distinguish the coun-
tries that are atypical or to motivate research to account for this. In places such
as Sri Lanka and the Indian state of Kerala, which have achieved better health
status than might have been expected given their aggregate income or rates of
poverty, this has often happened as a result of public actions that directly affect
health or nutrition (Anand and Ravallion 1993). Similarly, but less optimisti-
cally, in countries where nutritional status has improved less rapidly than might
have been expected given their income growth, this may indicate a need for spe-
cific investments in human resources (Alderman and Garcia 1994).
But most studies addressing the causal link between income growth and mal-
nutrition have focused on the response of nutrient consumption to changes in
income (Strauss and Thomas 1995; Bouis and Haddad 1992). Surprisingly,
there has been no systematic multicountry analysis of the causal relationship
between income and malnutrition. This article fills that gap. Our goal is to
answer this question: How far does moderately rapid income growth take us
toward reducing the rate of child malnutrition in line with the Millennium
Development Goal? We use an anthropometric measure--low weight for age--
of child nutritional status as an outcome of household decisions on health and
child care as well as on food consumption. We study the extent to which greater
resources at the household as well as the national level explain differences in
this crucial outcome.
Using household survey data from 12 countries as well as aggregate data on
a set of 61 developing economies, we model the relationship between child un-
derweight and per capita income, proxied by total household consumption per
capita in the micro studies and by per capita gross domestic product (gdp) esti-
mated using 1987 purchasing power parity (ppp) rates in the cross-country re-
gressions. We then use the model to predict the declines in malnutrition that can
be expected from a sustained 2.5 percent annual increase in per capita income
from the date of the survey (in the 1990s) to 2015. Even at this moderately rapid
growth rate, in 9 of 12 countries declines in malnutrition rates fall short of the
Millennium Development Goal target. We conclude that income growth can play
an important part in reducing malnutrition but that it is not enough. We suggest
Haddad and others 109
(but cannot prove in this study) that increasing the number and effectiveness of
direct nutrition interventions is crucial if nutrition goals are to be met.
I. Data Sets and Models
In this section we describe the two data sources used to derive estimates of the
response of child malnutrition to per capita income growth and outline the models
used to generate the results reported in the following section.
The Household Surveys
We investigate how household resources affect the nutritional status of preschool
children using household surveys from 12 countries.2 The countries were selected
from those with nationally representative household data for the 1990s to cover
a range of locations, spanning four continents. They differ appreciably in their
economic situation, including gdp per capita and national rates of malnutrition
(table 1).3 Even so, there is a common thread in the data: in all the countries
studied an integrated household survey was undertaken in the 1990s using a
multipurpose, modular, living standards survey following a format utilized in
more than 20 countries (Grosh and Glewwe 2000). These surveys collect data
on children's height and weight as well as information on total expenditures and
other socioeconomic conditions of the household.
The measure of nutritional status (N) that we study is weight for age, consid-
ered a general indicator of the nutritional status of populations (Alderman 2000;
who 1995). It is converted into standardized units called z-scores after compari-
son with the U.S. data chosen as an international reference by the World Health
Organization (who). The z-scores are derived after subtracting the age- and
gender-specific means from the reference data and after dividing by the corre-
sponding standard deviation. Like most of the literature, we pay particular at-
tention to the proportion of children below two standard deviations from the
median for the reference population. We refer to children with a weight-for-age
z-score of less than �2 as underweight. In the reference population, 2.3 percent
have z-scores of less than �2, and 16.0 percent have z-scores of less than �1. These
levels might be expected for a normal population and provide a basis for com-
parison. But because there is no sharp difference in risk of mortality or func-
tional impairment at this or any other commonly used cutoff level (Pelletier 1994),
the regressions focus on nutritional status, not the probability of malnutrition
defined in terms of a z-score of less than �2.
Countries with higher per capita income tend to have less malnutrition (see
table 1). But there are exceptions. Although South Africa has the highest income
2. The age range was usually 0�60 months. In Kenya the age range was 6�60 months, and in Nepal
0�36 months.
3. Because of data unavailability, we were unable to cover the half of the world's population that
lives in China and India.
Table 1. Summary of Household Survey Data Sets
Annual percentage Child malnutrition rate
Preschool Per capita change in per capita (percentage of preschool
children Year of Maternal gdp gdp (ppp)a children underweight)
included in sample height (U.S. dollars)
Country regressions survey covered 1998 1975�99 1990�99 Male Female All
Egypt, Arab Rep. 1213 1997 Yes 1290 2.9 2.4 10.3 11.1 10.7
Jamaica 752 1995 No 1680 0.1 �0.6 4.9 5.2 5.0
Kenya 7626 1994 No 330 0.4 �0.3 20.9 18.4 19.7
Kyrgyz Republic 1679 1997 Yesb 350 �5.3 �6.4 13.4 13.1 13.3
Morocco 1979 1990�91 Yes 1250 1.4 0.4 14.7 15.4 15.0
Mozambique 3268 1997 No 210 1.3 3.8 23.8 21.7 22.8
Nepal 1560 1996 No 210 1.8 2.3 50.4 45.6 48.1
Pakistan 3076 1991 Yes 480 2.9 1.3 48.4 43.2 45.7
Peru 3075 1997 No 2460 �0.8 3.2 7.5 5.5 6.5
Romania 3625 1994 No 1390 �0.5 �0.5 7.9 4.8 6.4
110 South Africa 4132 1993 No 2880 �0.8 �0.2 18.2 17.7 18.0
Vietnam 2637 1993 Yes 330 4.8 6.2 39.8 41.5 40.7
aData on real per capita gdp (adjusted for ppp) are from undp (2001).
bMother codes were not documented so this data could not be linked to children.
Source: Egypt: Integrated Household Survey (eihs) conducted under ifpri Food Security Research Project in Egypt, March�May 1997. Food Consumption and
Nutrition Division (fcnd), April 2000, Documentation of fcnd data sets collected between 1994 and 1999, p. 7. fcnd, International Food Policy Research Institute,
Washington, D.C. Jamaica: World Bank (2002), "Jamaica Survey of Living Conditions (jslc) 1998�2000--Basic Information," mimeo, Poverty and Human Re-
sources Division, Development Research Group, World Bank, Washington, D.C., www.worldbank.org/html/prdph/lsms/country/jm/docs/binfo2000.pdf. Kenya: Re-
public of Kenya (1996), "Welfare Monitoring Survey II--Basic Report," Central Bureau of Statistics, Nairobi, www4.worldbank.org/afr/poverty/pdf/docnav/00643.pdf.
Kyrgyz Republic: Kyrgyz Poverty Monitoring Survey 1997 (kpms), www.worldbank.org/lsms/country/kyrgyz/docs/kyrbif2.pdf. Morocco: Morocco Living Standards
Survey 1990/1 (mlss), www.worldbank.org/lsms/country/mo91/docs/mo91binf.pdf. Mozambique: Safety Net Design and Poverty Monitoring in Mozambique, Feb-
ruary 1996 through April 1997, fcnd, April 2000, Documentation of fcnd data sets collected between 1994 and 1999, p. 23. fcnd, International Food Policy
Research Institute, Washington, D.C. Nepal: World Bank (2002), "Nepal Living Standards Survey I 1995/96--Survey design and implementation," mimeo, Devel-
opment Research Group, World Bank, Washington, D.C., www.worldbank.org/lsms/country/nepal/nep96bidr.pdf. Peru: 1997 Encuesta Nacional de Hogares sobre
Medicion de Niveles de Vida (enniv) survey, collected by the Instituto Cuanto. For further details see annex 1 of World Bank (1999). Poverty and Social Develop-
ment in Peru, 1994�1997. Washington, D.C.: World Bank. Pakistan: World Bank (1985) "Basic Information--Pakistan Household Survey (pihs) 1991," mimeo,
Poverty and Human Resources Division, World Bank, Washington, D.C., www.worldbank.org/html/prdph/lsms/country/pk91/pk91.pdf. Romania: World Bank (1998),
"Basic Information--Romania Integrated Household Survey (rihs)," mimeo, Poverty and Human Resources, Development Research Group, World Bank, Washing-
ton, D.C. www.worldbank.org/lsms/country/romania/rm94bid.pdf. South Africa: South Africa Integrated Household Survey,. 1993/4, School of Economics, Univer-
sity of Cape Town, www.worldbank.org/html/prdph/lsms/country/za94/za94data.html. Vietnam: World Bank (2001), "Vietnam Living Standards Survey (vlss),
1997�98--Basic Information," mimeo, Poverty and Human Resources Division, World Bank, Washington, D.C., www.worldbank.org/lsms/country/vn98/vn98bif.pdf.
Haddad and others 111
in our sample of 12 countries, its malnutrition rates are little better than those in
Kenya, whose per capita income is less than an eighth of South Africa's. But our
focus with the household data is on the relationships between household resources
and nutritional outcomes across households within a given country. As is gener-
ally the case, we presume that expenditures reflect a household's long-run in-
come potential. Thus we estimate regressions for nutritional outcomes as a
function of the log of per capita household expenditures (Y).
Additional regressors include the education levels of the child's mother and
father (or, where parentage is unknown, a proxy).4 Beyond income earning ability,
education captures--though imperfectly--the ability of each parent to obtain
and use information about appropriate caring practices and health services for
the child. To account for different patterns of malnutrition by age, all the re-
gressions contain six dummy variables for age brackets. In addition, to control
for health- and sanitation-related correlates of income that may have an indepen-
dent effect on nutrition, the regressions include indicators for the type of drinking
water and toilet used.5 Moreover, in countries where there are significant ethnic
differences that relate to access to infrastructure--for example, Peru or South
Africa--the regressions also include dummy variables for ethnic background.6 The
height of the mother--an indicator of genetic endowment and of growth and de-
velopment in the womb--is included in the regressions when this information is
available. Finally, all models include demographic variables, such as household
size and the percentage of household which lies in different age groups.
We undertake two specifications of the model. Model 1 includes expenditures
but excludes health, water, and sanitation infrastructure both external and in-
ternal to the household.7 Model 2 controls for the infrastructure in the commu-
nity that is external to the household (E) by including cluster-level fixed effects.
4. If the child's father could not be identified, the education of the most educated adult male in the
household was used. In Jamaica and Kenya neither of a child's parents was identified, so the education
levels of the household head and his or her spouse were used instead. Education was typically measured
in years. For Kenya, however, for which this information was not available, dummy variables for edu-
cation level were used instead.
5. Typically the distinction was whether the household had piped drinking water within the dwell-
ing or not and whether it had a flush toilet (see Burger and Esrey 1995 for a discussion of the role of
water and sanitation interventions in reducing undernutrition).
6. However, who (1995) advocates using a single international reference for child growth. The rea-
son is that there are few if any ethnic differences in growth patterns of young children, and children
from privileged or middle-class families in developing economies generally have height and weight dis-
tributions that do not differ from international references.
7. For both the household survey and the cross-country regressions we log the per capita expendi-
ture variable to minimize the influence of extreme values of per capita expenditure. This also increases
the marginal effect of resources on nutrition at lower income levels, because the marginal effect is the
estimated coefficient on the log of expenditures divided by the observed level of expenditures. We con-
duct nonnested tests (Davidson and MacKinnon's J-test as outlined in Greene 2000) to determine the
appropriateness of this specification compared with a model linear in expenditures. In cases where the
test proved conclusive, the log model was favored in seven cases and the linear in two. In 3 of 12 cases
the test proved inconclusive.
112 the world bank economic review, vol. 17, no. 1
That is, the model includes a dummy variable for each sample cluster. This dummy
variable also picks up the effect of common attitudes and resources in the com-
munity or special local circumstances. In addition, model 2 includes the vari-
ables for infrastructure within the household (I) through access to piped water
and sanitation. The two models can be labeled as follows:
(1) N = N(Y)
(2) N = N(Y, E, I)
Model 2 can be considered to give the short-run effect of increasing household
income or consumption, holding external infrastructure and internal health in-
frastructure constant. Over a longer period a household whose income increases
may choose to invest in water and sanitation or may have such investments made
on its behalf by the public sector. Model 1, for which the short-run interpreta-
tion of the coefficient on income is biased to the degree that health and sanita-
tion effects that influence nutritional status are correlated with household income,
may better represent the total effect of resources in a long-run scenario.8
There are several reasons to suspect the endogeneity of the income variable in
both models. The most obvious reason is measurement error in income or in the
expenditure variable that we use in lieu of income. As is well known, if random
measurement error is present in an explanatory variable, ols estimates will be
biased toward zero. Another potential cause of endogeneity of income is time
allocation decisions that affect both income generation through labor supply and
child nutrition through child care. Consequently, we estimate the models using
both ols and instrumental variables, both with and without the community fixed
effects. Although there are differences in the nature and number of identifying
variables in each data set, we use land and livestock holdings as well as other
assets and durable goods in per capita terms, where available, as identifying in-
struments. In all cases we test the strength of our proposed identifying instru-
ments in predicting per capita expenditures (an F-test), whether it is valid to
exclude the proposed identifying instruments from the malnutrition equation (a
chi-squared test for overidentification), and the significance of the difference
between the consistent instrumental variables estimates on income and the effi-
cient ols estimates (a chi-squared Hausman test).9
The Cross-Country Data for 61 Countries, 1970�95
The dependent variable used in the cross-country analysis is the prevalence of
children under age five who are underweight for their age--that is, whose weight
8. In principle, the education coefficient of model 2 can be used to derive the effect of long-run in-
come growth on nutrition that is mediated by increased parental education that may also be driven by
income growth under any assumption of changes in education.
9. The list of instruments and the full set of results of these tests are available from the authors.
Further details on the tests are in Bound and others (1995) and Davidson and MacKinnon (1993).
Haddad and others 113
falls more than two standard deviations below the median for their age. All the
data for this variable are survey-based aggregates. Most of the data (75 percent)
are from the who's Global Database on Child Growth and Malnutrition (who
1997). These data have been subjected to strict quality control standards.10 The
rest of the data are from acc/scn (1993) and World Bank (1997), and we have
subjected these data to similar quality checks. We match each weight-for-age
survey year with the corresponding year's value of per capita gdp expressed in
1987 U.S. dollars adjusted for ppp. The gdp data are from the World Bank's
World Development Indicators 1998 (1998).11
The data set covers 61 developing economies, accounting for more than 80
percent of the developing world's population. Each country has at least two
observations, and many have three or four. The total number of country-year
observations is 175, spanning the period 1970�95 (Smith and Haddad 2000).12
II. Results: What Is the Impact of Income on Malnutrition?
In this section we present the regression results for the effects of income growth
at household and national levels on child malnutrition. We describe first the
results from the 12 household surveys and then the results from the cross-country
analysis.
Household Survey Results: Per Capita Household Income
and Child Malnutrition
Table 2 presents estimates of the coefficient of the logarithm of per capita con-
sumption (our proxy for per capita income) for models 1 and 2.13 It gives both
ols and instrumental variables estimates, with and without mother's height where
that variable is available. Several things are worth noting.
First, as expected, the log of per capita household consumption has a positive
relationship with the nutritional status of children as measured by weight for age
in all the countries studied. All the ols estimates of model 1 (without controls for
infrastructure) differ significantly from zero, as do most of the other estimates.
10. The criteria for inclusion in the who database are a clearly defined population-based sampling
frame, permitting inferences to be drawn about an entire population; a probabilistic sampling proce-
dure involving at least 400 children; use of appropriate equipment and standard measurement tech-
niques; and presentation of data in the form of z-scores in relation to the reference population chosen
by who (1997).
11. These gdp data are reported only for 1980 to the present. To arrive at comparable ppp gdp per
capita figures for the data points in the 1970s, it was necessary to impute growth rates from the data
series on gdp in constant local currency units and apply them to countries' 1987 ppp gdps.
12. Related work by Smith and Haddad (2002) indicates that instrumenting per capita gdp with
the investment share of gdp and the foreign investment share of gdp does not allow us to reject the
exogeneity of per capita gdp in the cross-country sample. Thus, we do not instrument per capita gdp
in the cross-country regressions.
13. Table A-1 presents these results in more detail and lists the instruments used. The full set of
results for each country is available from the authors.
Table 2. Summary of Estimates of the Effect of Per Capita Household Consumption on Weight
for Age of Preschool Children, Selected Developing Economies
Model 1: N = N(Y) Model 2: N = N(Y, E, I)
ols iv ols iv
ols with iv with without without ols with iv with without without
mother's mother's mother's mother's mother's mother's mother's mother's
height height height height height height height height
Egypt, Arab Rep.
Estimated coefficienta 0.1438 0.3600 0.1713 0.4007 0.1652 0.2977 0.1736 0.3176
t-statistic 2.09 2.00 2.47 2.21 1.98 1.30 2.07 1.38
Hausman testb p = 0.1948 p = 0.1698 p = 0.5360 p = 0.5029
Jamaica
Estimated coefficient 0.257 0.742 0.191 0.411
114 t-statistic 3.13 3.10 2.11 1.51
Hausman testb p = 0.027 p = 0.393
Kenya
Estimated coefficienta 0.137 0.499 0.142 0.417
t-statistic 8.02 7.38 6.36 4.64
Hausman testb p = 0.000 p = 0.01
Kyrgyz Republic
Estimated coefficienta 0.2157 0.2893 0.1619 0.3553
t-statistic 3.48 1.68 2.19 1.81
Hausman testb p = 0.6469 p = 0.2882
Morocco
Estimated coefficienta 0.4274 0.7174 0.4857 0.7814 0.1879 0.6007 0.2333 0.6330
t-statistic 8.44 9.18 9.62 10.16 2.78 3.86 3.46 4.10
Hausman testb p = 1.12e�06 p = 3.55e�07 p = 0.0032 p = 0.0040
Mozambique
Estimated coefficienta 0.3127 0.4595 0.1860 0.3403
t-statistic 10.68 8.76 3.94 3.62
Hausman testb p = 0.000746 p = 0.05807
Nepal
Estimated coefficienta 0.319 0.971 0.204 0.533
t-statistic 6.16 5.15 2.98 2.78
Hausman testb p = 0.00 p = 0.068
Pakistan
Estimated coefficienta 0.231 0.471 0.240 0.478 0.075 0.400 0.085 0.405
t-statistic 4.77 3.29 4.96 3.36 1.34 2.25 1.52 2.28
Hausman testb p = 0.073 p = 0.073 p = 0.053 p = 0.056
Peru
Estimated coefficienta 0.2504 1.2001 0.2056 0.8150
t-statistic 5.51 5.38 4.09 3.52
Hausman testb p = 0.0000139 p = 0.0.0069
Romania
Estimated coefficienta 0.140 0.180 0.287 0.658
115 t-statistic 3.28 2.00 2.78 2.89
Hausman testb p = 0.279 p = 0.066
South Africa
Estimated coefficienta 0.2089 0.2790 0.1780 0.0807
t-statistic 5.39 1.48 3.45 0.28
Hausman testb p = 0.7048 p = 0.7327
Vietnam
Estimated coefficienta 0.265 0.437 0.293 0.471 0.198 0.261 0.105 0.275
t-statistic 6.73 7.02 7.37 7.52 1.76 2.55 1.87 2.67
Hausman testb p = 0.000 p = 0.000 p = 0.057 p = 0.049
Note: The table shows results for the log of per capita expenditure (lnpcxp). Estimates are used in the projections in table 3.
IV is instrumental variables.
alnpcxp.
bOLS vs. IV (chi squared).
Source: Authors' calculations.
116 the world bank economic review, vol. 17, no. 1
Second, the estimated coefficients on the log of per capita consumption are
usually larger in model 1 than in model 2. The exceptions to this are the Arab
Republic of Egypt and Romania. The general pattern is consistent with the inter-
pretation that model 1 captures the long-run effect of income on malnutrition.
Third, the instrumental variables estimates are, without exception, larger than
the ols estimates. The differences range from 29 percent in Romania to 500
percent in Peru. These differences are consistent with a high degree of measure-
ment error on the per capita consumption variable.
Fourth, the instrumental variables estimates differ significantly from zero and
differ significantly at the 5 percent level from the ols estimates for 8 of the 12
countries. ols estimates are preferred to the instrumental variables estimates for
3 of the 12 countries. For the Kyrgyz Republic and South Africa we cannot gen-
erate significant instrumental variables estimates for either model 1 or 2. For
Romania instrumental variables estimates can be generated that differ signifi-
cantly from zero, but the Hausman test fails to reject the equality of ols and
instrumental variables estimates even at the low threshold of 20 percent, arbi-
trarily selected to take into account the low power of the test. For the remaining
country, Egypt, we selected the instrumental variables estimate (0.36) rather than
the lower ols estimate (0.14) for the subsequent projections even though the
Hausman test only rejected the equality of the estimates at the 19 percent level.
Fifth, the estimated coefficients on the log of per capita consumption are larger
in the absence of mother's height. The differences (in our preferred specifica-
tions) range from 1 percent in Pakistan to 11 percent in Egypt. These differences
are consistent with the hypothesis that failing to control for mother's height will
lead to a bias due to omitted variables (Alderman 2000). The bias appears mod-
est in the four cases in which we can test for this, however.
Sixth, if we focus on our preferred estimates of model 1 (table 2), the mean
coefficient is 0.54--implying that doubling household income will increase weight
for age by half a standard deviation from the median for the reference popula-
tion. The median coefficient is 0.47. But the coefficients vary widely across coun-
tries, from 0.14 for Romania to 1.20 for Peru.
The results reported in table 2 are based on regressions that have nutritional
status as a dependent variable. Though this approach uses more information in
the data sets than one focusing on the probability of crossing a threshold, it does
not allow us to directly infer the effect of income growth on malnutrition rates.
Under the assumption of a neutral distribution of income growth, however, it is
relatively straightforward to simulate expected change in the prevalence of mal-
nutrition between the year of a survey and 2015 (the reference point for the
Millennium Development Goals) using the coefficients in table 2.
Table 3 shows the expected proportional reduction in malnutrition after sus-
tained per capita income growth of 2.5 percent a year, using the estimates in
table 2 (all from model 1, the long-run specification). Because we force income
growth to be the same across countries, any differences in the effect of this growth
reflect the size of the estimated coefficient on income and the density of the dis-
Haddad and others 117
Table 3. Projected Child Malnutrition Rate with 2.5 Percent Annual Growth
in Per Capita Income from the 1990s to 2015, Selected Developing Economies
Child Projected child
Estimated malnutrition rate malnutrition
coefficient on in survey year rate in 2015 Change
log of per (percentage of (percentage of in child
capita preschool preschool malnutrition
expenditure children children rate Arc
Country from model 1 underweight) underweight) (percent) elasticity
Egypt, Arab Rep. 0.3600a 10.80 8.00 �25.95 �0.464
Jamaica 0.7415a 5.05 2.26 �55.26 �0.865
Kenya 0.4994a 19.63 11.38 �42.02 �0.618
Kyrgyz Republic 0.2157b 13.28 11.44 �13.90 �0.248
Morocco 0.7174a 13.79 6.11 �55.68 �0.670
Mozambique 0.4595a 23.04 16.43 �28.69 �0.513
Nepal 0.9710a 48.08 25.99 �45.94 �0.767
Pakistan 0.4705a 45.73 34.67 �24.18 �0.299
Peru 1.2001a 7.32 2.70 �63.11 �1.127
Romania 0.1396b 6.40 5.54 �13.36 �0.197
South Africa 0.2089b 18.02 15.54 �13.79 �0.191
Vietnam 0.4372a 40.65 28.13 �30.78 �0.427
aInstrumental variables estimate.
bols estimate.
Source: See table 1.
tribution of the nutritional status of the population slightly below the cutoff for
malnutrition at a z-score of �2. The assumed growth rate for per capita income
is relatively optimistic. Only 3 of the 12 countries achieved this growth rate over
the 1990s, although another 2 came close (see table 1). Over the 25-year period
ending in 1999, again only three countries achieved 2.5 percent per capita growth.
The cross-country data set confirms that the income growth rates used in our
simulations are optimistic. Based on all observations available (61 countries, 175
observations), the mean growth in per capita gdp between the earliest and lat-
est years for each country averages just 1 percent a year. In the countries for
which we have observations for all three decades, growth averaged only 0.65
percent a year.
For only 3 of the 12 countries--Jamaica, Morocco, and Peru--does per capita
income growth of 2.5 percent result in a halving of the malnutrition rate by 2015.
Among the 12 countries, these 3 rank first, third, and sixth, respectively, by lowest
initial rate of malnutrition, although there is no statistically significant correla-
tion between the initial malnutrition rate and the projected decline across the 12
countries. The relative decline ranges from 13 percent in Romania to 63 percent
in Peru, averaging 34 percent (the median decline is 30 percent).
These projected declines are likely to be on the high end for several reasons.
First, by using estimates from model 1, we assume that as a household's income
improves, so does the health and sanitation infrastructure to which the household
118 the world bank economic review, vol. 17, no. 1
has access, both internally and externally. If we assume that infrastructure and
community fixed factors do not improve (basing our estimates on model 2), sus-
tained growth of 2.5 percent would reduce malnutrition by an average 27.4 per-
cent by 2015.14 Second, we assume that every household experiences the same rate
of income growth, an assumption that forces growth to be broadly based. Third,
we assume fairly robust growth of per capita income. If we assume a more modest
rate of, for example, 1.25 percent a year (achieved by only half the 12 countries in
1990�99), none of the 12 countries would meet the target of halving malnutrition
rates by 2015. Fourth, by using the estimated coefficients from the log specifica-
tion on per capita consumption, regardless of what the nonnested tests conclude,
we force the estimated effect of income on nutrition to be relatively large for poorer
households (which tend to contain proportionately more underweight children).
Before looking at the effect of gdp growth on cross-country regressions, we
discuss the coefficients of the auxiliary variables included in the household re-
gressions to reduce the bias due to missing variables, such as parental education
and the infrastructure terms, focusing our attention on model 2.
Parental characteristics are often important determinants of anthropometric
status (table 4). This is particularly true for mother's height, which had a posi-
tive and significant relationship with the child's nutrition in all the countries for
which this information was available. Years of parental education are positive
and significant determinants of anthropometric status in just over a third of all
cases. The lack of significance may be surprising given the conventional wisdom,
although it mirrors the findings of Sahn and others (1999) based on Demographic
and Health Surveys for nine African countries.15 The estimates of the coefficients
are almost always positive and, taken together, make it unlikely that their true
value is zero. On average, an extra year of maternal education raises z-scores by
around 1.3 percent of a standard deviation of nutritional status. Paternal edu-
cation generally has a somewhat smaller effect (averaging 0.7 percent of a stan-
dard deviation), though it varies by country. On average, giving mothers and
fathers an extra six years of schooling each would raise weight for age by 12
percent of a standard deviation. Compare this with the 54 percent average change
predicted from doubling income.
In all cases the age bracket variables for the child were jointly significant and in
most cases individually so. The anthropometric data show no evidence of bias
against girls, even in countries where it is commonly suspected, such as Nepal and
Pakistan (see also Harriss 1995). z-Scores are almost always higher on average for
girls than for boys, although the differences are often statistically insignificant.
14. With estimated coefficients from model 2, the malnutrition rate would decline by 16.03 percent
in Egypt, 15.79 percent in Jamaica, 36.10 percent in Kenya, 11.66 percent in the Kyrgyz Republic,
49.81 percent in Morocco, 22.70 percent in Mozambique, 27.20 percent in Nepal, 19.11 percent in
Pakistan, 45.33 percent in Peru, 58.19 percent in Romania, 7.76 percent in South Africa, and 19.56
percent in Vietnam.
15. In one specification parental education variables were significant determinants of height for age
in only 11 of 32 cases studied by Sahn and others (1999, table 14A).
Haddad and others 119
Table 4. Coefficients on Parental Characteristics, Selected
Developing Economies
Country Father's education Mother's education Mother's height
Egypt, Arab Rep. �0.0106 (1.29) 0.0019 (0.20) 0.0240 (3.50)
Egypt, Arab Rep. �0.01049 (1.27) 0.0033 (0.34)
Jamaica 0.0052 (0.24) 0.0165 (1.15) n.a.
Kenya 0.0016 (0.35) 0.0144 (3.77) n.a.
Kyrgyz Republic 0.0024 (0.14) 0.0580 (2.99) n.a.
Morocco 0.0006 (0.01) �0.0358 (0.15) 0.0270 (4.97)
Morocco 0.0076 (0.13) �0.038 (0.16)
Mozambique 0.0023 (0.28) 0.0261 (2.24) n.a.
Nepal 0.0212 (2.76) 0.0146 (1.20) n.a.
Pakistan 0.0198 (2.68) 0.0311 (2.79) 0.0060 (2.38)
Pakistan 0.0218 (2.97) 0.0308 (2.76)
Peru �0.0165 (2.53) 0.0284 (3.47) n.a.
Romania 0.0480 (2.63) �0.0185 (�0.89) n.a.
South Africa 0.0167 (1.48) 0.0049 (0.62) n.a.
Vietnam �0.0042 (�0.53) 0.0182 (2.10) 0.0253 (6.15)
Vietnam �0.0048 (�0.61) 0.0190 (2.17)
Note: The dependent variable is weight for age (z-score) preschool children. The
coefficients are ols estimates from model 2. n.a., Not available.
Source: Authors' calculations.
Cross-Country Results: Per Capita GDP and Child Malnutrition
Table 5 presents the mean prevalence of malnutrition in our cross-country sample,
both for all the countries and for the subsample for which we have observations
in each decade. We report both unweighted cross-country means and means
weighted by country population. Comparisons of trends in malnutrition rates
over time are complicated by our lack of observations for China in the 1970s
and India in the 1980s. But the data do illustrate the cross-sectional variation of
malnutrition with national income.
Figure 1 plots the predicted negative relationship between smoothed malnu-
trition rates and per capita gdp based on the smoothed regression routine for
each decade. The association between gdp and nutrition has been fairly con-
stant; the line for the 1970s runs parallel to those for the next two decades. At
any given level of gdp in the 1980s or 1990s, a country could expect a lower
rate of malnutrition than in the 1970s. That is, even in countries with stagnant
economies, the expected rate of malnutrition in the 1980s was lower than that
in the 1970s. Plausible candidates that may account for this change between the
1970s and 1980s include improvements in technology that are not strongly re-
lated to income or investment in the countries in the sample, such as the promo-
tion of oral rehydration salts and mass immunization. In addition, the average
price of food was higher in the 1970s. Though it is also true that the average
education of women (as well as men) improved in the period, this is less likely to
be an explanation, because (as will be discussed) the 1970s imply higher malnu-
120 the world bank economic review, vol. 17, no. 1
Table 5. Mean Child Malnutrition Rate in Cross-Country Data
Mean child malnutrition rate
(percentage of preschool children underweight)
Decade Unweighted Population weighted Observations
All countries
1970s 29.18 50.8 30
1980s 24.23 29.0 74
1990s 23.80 28.5 71
All 24.90 175
Countries with observations in all decades
1970s 27.07 33.9 18
1980s 20.69 26 27
1990s 19.65 24.5 22
All 22.06 67
Source: who (1997).
trition even in regressions that control for education. Moreover, the improve-
ment in education continued and indeed accelerated in many countries into the
1990s, but the curve for that decade is not appreciably below that for the 1980s.
Table 6 reports models of malnutrition rates as a function of the log of per
capita gdp, female secondary school enrollment, access to safe water, and de-
cade dummy variables. The models do not explore how different structures of
gdp growth influence malnutrition, although the fixed effects results do have
some control for such differences. The ols results (without the variable for ac-
cess to safe water) are analogous to model 1 (column 1 of table 6), and the country
fixed effects estimates (with the variable for access to safe water) are analogous
to model 2 (column 2).16
The decline in malnutrition rates over time suggested in figure 1 is confirmed
by the negative signs of the dummy variables for the 1980s and 1990s (signifi-
cant at 5 percent for the 1980s only in the model 2 specification) relative to the
1970s. Model 1 estimates indicate a negative and significant effect of per capita
gdp on malnutrition rates. By dividing the coefficient of the logarithm by the
16. Although Pritchett and Summers (1996) present evidence that gdp can be treated as exogenous
in cross-country health regressions, we explored potential concerns about measurement error in the
explanatory variables using a procedure suggested by Griliches and Hausman (1986). For the 36 coun-
tries with more than two observations, we generated two sets of fixed-effects estimates by differencing
out the fixed effects in two different ways. First, we differenced observations t1 and t2; second, we
differenced the first and last observations. The two sets of estimates were similar, especially for log per
capita gdp (�6.13, t = 1.29 in the first case, and �6.91, t = 1.91 in the second). Because attenuation bias
does not worsen appreciably with shorter periods between observations, we conclude that measure-
ment error in the explanatory variables does no major violence to our findings on the size of the esti-
mated coefficient on the log of per capita gdp (Johnston and DiNardo 1997). This approach also partially
addresses a concern about the education variables that are generally less useful in time series of aggre-
gate data than in household data (Krueger and Lindhal 1999), though the longer the interval, the less
of a concern this is.
Haddad and others 121
Figure 1. Fitted Relationship between Child Malnutrition Rate and Per
Capita gdp in Developing Economies, 1970s, 1980s, and 1990s
45
40
35
are
30
children
who
of 1970's
25
five
underweight 20 1980's
under
Percentage 15 1990's
10
5
0
0 1000 2000 3000 4000 5000 6000 7000 8000 9000 10000
Per capita GDP
(1987 U.S. dollars adjusted for purchasing power parity)
Note: The ksm command in Stata (V7) with only per capita GDP as an explanatory variable was used
to generate the smoothed curve. A bandwidth of 0.8 was used.
Source: Authors' calculations.
mean rate of malnutrition in the sample countries reported in table 5, we derive
an elasticity at the mean of �0.51, comparable to the mean (�0.53) of the arc
elasticities reported in table 3 from the survey-based estimates. As expected, the
inclusion of fixed effects and the variable for access to safe water in model 2
leads to a smaller estimate of the effect of income growth. In column 2 of table
6 the coefficient on per capita gdp drops to 59 percent of its value in column 1.
This general result holds for fixed effects estimation with the variable for access
to safe water and without it (not reported here). It suggests that there are many
time-invariant unobservable factors that are positively associated with both high
(low) income and low (high) malnutrition, biasing the ols estimates upward.
The estimated coefficient on the log of per capita gdp in column 2 of table 6
implies that 2.5 percent annual growth in per capita gdp between 1995 and 2015
would reduce the malnutrition rate by 8 percentage points, or 32 percent of the
initial rate (compared with 34 percent, the mean relative decline for the 12 sur-
vey countries). The results refute a hypothesis that per capita gdp growth fails
to improve the nutritional status of the most vulnerable. This improvement in
nutrition related to gdp growth may be a direct effect of economic growth on
the income of households with malnourished members (presumably the poor)
or an indirect effect of this growth on the infrastructure of the country--or a
combination of the two.
122 the world bank economic review, vol. 17, no. 1
Table 6. ols and Country Fixed Effects Regressions,
Cross-Country Data
Explanatory variable (1) ols (2) Country fixed effects
Log of per capita gdp �12.673 �7.44
(8.00)** (2.89)**
Female secondary school enrollment �0.011 �0.088
(0.19) (1.13)
Percentage of households with access �0.055
to safe water (1.18)
Decade = 1980s �4.411 �4.07
(1.77) (2.66)*
Decade = 1990s �6.385 �4.18
(2.52)* (2.19)*
Constant 124.220 89.80
(11.24)** (4.92)**
Observations 175 175
Countries 61 61
R2 0.45 0.43
*Significant at the 5 percent level.
**Significant at the 1 percent level.
Note: The dependent variable is the prevalence of preschool children who are under-
weight for their age (z-score less than �2). The numbers in parentheses are the absolute value
of t-statistics.
Source: Authors' calculations.
The percentage reductions in malnutrition rates estimated using the survey
data are remarkably similar to those estimated using the cross-country data. Of
course, there is no automatic correspondence between the household regressions
and the cross-country results. For one thing, income growth rates estimated using
the national accounts data in the cross-country regressions do not closely track
those estimated using survey data on household expenditures (Deaton 2001). In
addition, the rate of income growth for the households at risk of malnutrition
may differ from the national average, depending on whether inequality is increas-
ing or declining. Moreover, the cross-country results might be biased downward
because of mismeasurement in ppp. Conversely, one might expect the cross-
country results to give higher income elasticities than those based on household
survey data, because the second are conditioned on time-varying as well as time-
invariant country-level factors. For example, if all households in a survey are
subject to the same national health system, household-level estimates of income
effects will not include the indirect effects on the performance of the system
from rising national income. Thus, it is reassuring that our main results on the
expected effect of income growth are fairly robust to the alternative source of
income data.
Our cross-country estimates have not yet explicitly addressed income distri-
bution. This omission is important for two reasons. First, it is plausible that the
inequality in a country affects the allocation of resources to basic health and
Haddad and others 123
similar services. Second, for our cross-country model to be consistent with the
semi-logarithmic specification at the household level, we need to accommodate
the fact that the per capita gdp variable is not equivalent to the average of the
logarithm of income. We cannot re-create that average with the aggregate data
available. However, the misspecification of the income variable when the true
model is semi-logarithmic is explicitly related to Theil's inequality measure. This,
too, is unavailable with the aggregate data, but a related measure is found in the
Gini coefficients in the Deininger and Squire Data Set on income inequality (World
Bank 2002). Although this is not the perfect correction for using per capita gdp
in a model based on a logarithmic income response, it serves as a conditioning
variable to reduce any error in the per capita gdp variable, though it does so
imprecisely. Because the Gini coefficient variable picks up the aggregation bias
as well as the possible causal relationship between inequality and the effect of
income growth on nutrition, there is no clear expectation for the sign.
From the Deininger and Squire Data Set it is clear that inequality measures
change over time and thus are not adequately controlled for in the fixed-effects
estimates. Merging that data set's self-declared "high-quality" data on the Gini
coefficient by country and year into our data set reduces the number of observa-
tions from 175 to 96 and the number of usable observations (those for countries
with more than one observation) to 79 (or 31 countries). Table 7 presents re-
gressions similar to those in table 6--but on this much smaller data set--both
with and without the Gini coefficient variable. This variable does not differ sig-
nificantly from zero at the 5 percent level in either the ols or the country fixed
effects specification. But it does have a negative coefficient. Importantly, intro-
ducing the Gini coefficient does not substantially alter the size of the estimated
coefficient on the log of per capita gdp.
III. Conclusions
Both the cross-country and the household-level results show that sustained in-
come growth could lead to a sizable reduction in malnutrition in the next de-
cade or so. Even with no change in community and household infrastructure,
rates of malnutrition (low weight for age) are projected to decline by an average
of around 27 percent by 2015 if countries can achieve per capita income growth
of 2.5 percent a year. Allowing community and household infrastructure to
change over time increases the effect of the growth to a 34 percent reduction in
national malnutrition rates. Cross-country regressions imply similar reductions.
The cross-country estimates add another dimension, showing that historical
patterns of income distribution are consistent with income growth leading to
marked improvements in nutrition.
Although these results are encouraging, others are disturbing. First, only 3 of
the 12 countries sustained per capita income growth of more than 2.5 percent a
year in the 1990s. Second, even if all 12 countries had 2.5 percent growth over
the approximately 20-year period ending in 2015, only 3 would meet the target
124 the world bank economic review, vol. 17, no. 1
Table 7. ols and Country Fixed Effects Regressions with Gini Coefficient,
Cross-Country Data
Country
fixed
ols effects
including Country including
Gini fixed coefficient
Explanatory variable ols coefficient effects Gini
Log of per capita gdp �17.216 �15.196 �10.165 �9.242
(7.35)** (5.53)** (2.09)* (1.94)
Female secondary school enrollment �0.038 �0.039 0.027 �0.005
(0.44) (0.47) (0.25) (0.05)
Percentage of households with access �0.119 �0.146
to safe water (1.80) (2.20)*
Decade = 1980s �6.302 �7.025 �4.368 �4.159
(1.71) (1.89) (1.93) (1.88)
Decade = 1990s �10.230 �11.143 �4.494 �3.975
(2.60)* (2.81)** (1.57) (1.41)
Gini coefficient �0.302 �0.342
(1.38) (1.80)
Constant 165.400 163.766 113.283 123.911
(9.74)** (9.67)** (3.20)** (3.54)**
Observations 79 79 79 79
Countries 31 31 31 31
R2 0.56 0.57 0.54 0.56
*Significant at the 5 percent level.
**Significant at the 1 percent level.
Note: The dependent variable is the prevalence of preschool children who are underweight for their
age (z-score less than �2). The numbers in parentheses are the absolute value of t-statistics. The table
includes only countries with more than one observation for the Gini coefficient.
Source: Authors' calculations.
of reducing malnutrition rates by half. Third, among the countries that will not
meet that target even with sustained growth of 2.5 percent a year are those with
the highest current malnutrition rates--Nepal, Pakistan, and Vietnam. Fourth,
even if all economies managed to grow at a pace that would halve malnutrition
rates by 2015, each year a different cohort of preschool children--particularly
those under 36 months of age--would be irreversibly harmed.17 Do we need to
wait this long for malnutrition rates to be halved?
Though income growth can take us a long way toward meeting the target for
malnutrition, it is unlikely by itself to ensure that outcome. What will it take to
meet this target--and at a more rapid pace? There are many effective nutrition
and health interventions that could accelerate reductions in malnutrition in the
short run (Allen and Gillespie 2001). Some of these interventions--particularly
vitamin A supplementation for children under age five, iron supplementation for
17. Moreover, even if the target is met in countries with high initial malnutrition rates, this is no
cause for complacency: these countries will still be home to many undernourished preschool children.
Haddad and others 125
pregnant women, and some types of nutrition education and behavior change
initiatives--are more cost-effective than others (Gillespie and Haddad 2001).
Impact evaluations and other project-level assessments have shown that such
instruments are effective. The long-run income estimates based on the survey
data allow for improvements in health-related infrastructure, but only at a "busi-
ness as usual" rate. Unfortunately, because of data constraints, it is impossible
to compare the cost-effectiveness of current health infrastructure captured by
the surveys with that of the "best practice" set of nutrition interventions, espe-
cially when the health infrastructure is broadly defined and can fall within other
sectors, such as education, infrastructure, and agriculture.
Income growth is also part of this balanced strategy. Sustained per capita in-
come growth will go a long way toward halving child malnutrition rates by 2015.
Indeed, in the absence of income growth, the effect of direct nutrition interven-
tions is likely to be hampered despite their potential.
Even so, we can echo the conclusions of Berg (1981) and Reutlinger and
Selowsky (1976), who note that malnutrition would persist in the face of rapid
income growth in the absence of additional measures to address malnutrition
directly, whatever those measures might be. Our results point to the crucial
importance of pursuing a balanced strategy to accelerate reductions in malnu-
trition, though by themselves the results do not identify which investments are
more effective in which environment (see Gillespie and others 1996, for example).
Appendix Table A-1. Full Results on the Effect of Per Capita Consumption, Selected Developing Economies
Model 1: N = N(Y) Model 2: N = N(Y, E, I)
ols with iv with ols without iv without ols with iv with ols without iv without
mother's mother's mother's mother's mother's mother's mother's mother's
height height height height height height height height
Egypt, Arab Rep.
Estimated coefficienta 0.1438 0.3600 0.1713 0.4007 0.1652 0.2977 0.1736 0.3176
t-statistic 2.09 2.00 2.47 2.21 1.98 1.30 2.07 1.38
F-test on significance F(10,1188) = 20.45 F(10,1189) = 20.72 F(10,1061) = 16.17 F(10,1062) = 16.24
of identifying instruments (p = 0) (p = 0) (p = 0) (p = 0)
Overidentification testb 6.31 (df = 9) (pass) 6.31 (df = 9) (pass) 3.158 (df = 9) (pass) 3.518 (df = 9) (pass)
Hausman test, ols vs. ivb p = 0.1948 p = 0.1698 p = 0.5360 p = 0.5029
Instruments (10) Per capita values of animals owned (and � rural-urban dummy variable), acres owned (and � rural-urban
dummy variable), other savings and bank deposits, other property not in use, durable goods, household
126 enterprise, and agricultural machinery (tractors, threshers) (and � rural-urban dummy variable)
Jamaica
Estimated coefficienta 0.257 0.742 0.191 0.411
t-statistic 3.13 3.10 2.11 1.51
Relevance test F(6,730) = 17.02 (p = 0) F(6,716) = 14.53 (p = 0)
Overidentification testb 0.752 (df = 6) (pass) 0.075 (df = 6) (pass)
Hausman test, ols vs. ivb p = 0.027 p = 0.393
Instruments (5) Log per capita value of durable goods, log per capita unearned income, log per capita rooms,
1--receive food stamps, 1--applied for food stamps, 1--own house
Kenya
Estimated coefficienta 0.137 0.499 0.142 0.417
t-statistic 8.02 7.38 6.36 4.64
Relevance test F(6,7603) = 92.29 (p = 0) F(6,6481) = 73.24 (p = 0)
Overidentification testb 6.01 (df = 5) (pass) 1.53 (df = 5) (pass)
Hausman test, ols vs. ivb p = 0.00 p = 0.01
Instruments (6) Log per capita cattle, 1--no cattle, log per capita number of rooms in house, 1--household head is
commercial farmer, 1--household head is in business, 1--iron roof
Kyrgyz Republic
Estimated coefficienta 0.2157 0.2893 0.1619 0.3553
t-statistic 3.48 1.68 2.19 1.81
F-test on significance of F(6,1657) = 41.0 (p = 0) F(6,1602) = 44.13 (p = 0)
identifying instruments
Overidentification testb 2.351 (df = 5) (pass) 0.672 (df = 5) (pass)
Hausman test, ols vs. ivb p = 0.6469 p = 0.2882
Instruments (6) Per capita values of durable goods, livestock, business owned, housing and properties owned, other
assets and savings, and land
Morocco
Estimated coefficienta 0.4274 0.7174 0.4857 0.7814 0.1879 0.6007 0.2333 0.6330
t-statistic 8.44 9.18 9.62 10.16 2.78 3.86 3.46 4.10
F-test on significance of F(5,1956) = 291.38 F(5,1957) = 304.76 F(5,1814) = 86.72 F(5,1815) = 87.93
identifying instruments (p = 0) (p = 0) (p = 0) (p = 0)
Overidentification testb 7.718 (df = 4) (pass) 8.1139 (df = 4) (pass) 4.35 (df = 4) (pass) 2.97 (df = 4) (pass)
Hausman test, ols vs. ivb p = 1.12e�06 p = 3.55e�07 p = 0.0032 p = 0.0040
Instruments (5) 1--own cooker, 1--own refrigerator, 1--own stove with gas, 1--own color TV, 1--own black and
127 white TV
Mozambique
Estimated coefficienta 0.3127 0.4595 0.1860 0.3403
t-statistic 10.68 8.76 3.94 3.62
F-test on significance of F(16,3279) = 92.26 F(16,2513) = 53.20
identifying instruments (p = 0) (p = 0)
Overidentification testb 19.28 (df = 15) (pass) 21.90 (df = 15) (pass)
Hausman test, ols vs. ivb p = 0.000746 p = 0.05807
Instruments (16) Per capita land area (ha), per capita livestock value, 1--own refrigerator, 1--own fan, 1--own sewing
machine, 1--own loom, 1--own iron, 1--own radio, 1--own TV, 1--own color TV, 1--own air
conditioner, 1�own clock, 1�own telephone, 1--own car, 1--own motor bike, 1--own bicycle
Nepal
Estimated coefficienta 0.319 0.971 0.204 0.533
t-statistic 6.16 5.15 2.98 2.78
Relevance test F(6,1539) = 23.22 (p = 0) F(6,1539) = 30.32 (p = 0)
Overidentification testb 5.14 (df = 5) (pass) 9.04 (df = 5) (pass)
Hausman test, ols vs. ivb p = 0.00 p = 0.068
Instruments (6) Log value of consumer durables, log per capita land value, log per capita livestock value, log per capita
value of farm enterprise assets, log per capita value of nonfarm enterprise, 1--electric lighting
(continued)
Table A-1. (continued)
Model 1: N = N(Y) Model 2: N = N(Y, E, I)
ols with iv with ols without iv without ols with iv with ols without iv without
mother's mother's mother's mother's mother's mother's mother's mother's
height height height height height height height height
Pakistan
Estimated coefficienta 0.231 0.471 0.240 0.478 0.075 0.400 0.085 0.405
t-statistic 4.77 3.29 4.96 3.36 1.34 2.25 1.52 2.28
Relevance test F(6,3051) = 66.61 F(8,3052) = 62.26 F(6,2759) = 51.26 F(6,2760) = 51.34
(p = 0) (p = 0) (p = 0) (p = 0)
Overidentification testb 8.613 (df = 6) (pass) 8.305 (df = 6) (pass) 0.615 (df = 6) (pass) 0.308 (df = 6) (pass)
Hausman test, ols vs. ivb p = 0.073 p = 0.073 p = 0.053 p = 0.056
Instruments (6) Log per capita land in ha, log per capita rooms, 1--mud floor, 1--iron roof, 1--no land, 1--missing
data on housing
128 Peru
Estimated coefficienta 0.2504 1.2001 0.2056 0.8150
t-statistic 5.51 5.38 4.09 3.52
F-test on significance of F(4,3055) = 66.23 F(4,2680) = 67.22
identifying instruments (p = 0) (p = 0)
Overidentification testb 0.0001 (df = 3) (pass) 0.0001 (df = 3) (pass)
Hausman test, ols vs. ivb p = 0.0000139 p = 0.0.0069
Instruments (4) Per capita values of durable goods (and squared term) and house (and squared term)
Romania
Estimated coefficienta 0.140 0.180 0.287 0.658
t-statistic 3.28 2.00 2.78 2.89
Relevance test F(10,3597) = 107.95 F(10,1216) = 31.97
(p = 0) (p = 0)
Overidentification testb 13.77 (df = 9) (pass) 2.175 (df = 9) (pass)
Hausman test, ols vs. ivb p = 0.279 p = 0.066
Instruments (10) Log per capita value of consumer durables; log per capita value of domestic currency savings; wage
earners as proportion of household size; log per capita own house value; log per capita private rent;
log per capita public rent; dummy variables for private renting, for public renting, and for missing
monetary information on housing
South Africa
Estimated coefficienta 0.2089 0.2790 0.1780 0.0807
t-statistic 5.39 1.48 3.45 0.28
F-test on significance of F(5,4108) = 36.02 (p = 0) F(5,3755) = 24.67 (p = 0)
identifying instruments (p = 0) (p = 0)
Overidentification testb 5.372 (df = 4) (pass) 0.0001 (df = 4) (pass)
Hausman test, ols vs. ivb p = 0.7048 p = 0.7327
Instruments (5) Wage earners per household, per capita land owned (hectares), per capita value of vehicle, per capita
value of other machinery (such as motorized pumps), per capita value of other immovable assets
(such as land not in use)
Vietnam
Estimated coefficienta 0.265 0.437 0.293 0.471 0.098 0.261 0.105 0.275
t-statistic 6.73 7.02 7.37 7.52 1.76 2.55 1.87 2.67
Relevance test F(4,2616) = 441.64 F(4,2617) = 447.10 F(4,2317) = 244.98 F(4,2318) = 245.35
(p = 0) (p = 0) (p = 0) (p = 0)
Overidentification testb 7.120 (df = 4) (pass) 7.120 (df = 4) (pass) 0.791 (df = 4) (pass) 0.791 (df = 4) (pass)
129 Hausman test, ols vs. ivb p = 0.000 p = 0.000 p = 0.057 p = 0.049
Instruments (4) Log per capita value of durable goods, log per capita land in ha, log per capita value of livestock,
1--no land
Note: The table shows results for the log of per capita expenditure (lnpcxp). The dependent variable is weight for age (z-score) of preschool
children. Estimates are used in the projections in table 3. IV is instrumental variables.
alnpcxp.
bChi-squared.
Source: Authors' calculations.
130 the world bank economic review, vol. 17, no. 1
References
acc/scn (United Nations Administrative Committee on Coordination/Sub-Committee
on Nutrition). 1993. Second Report on the World Nutrition Situation. Vol. 2, Coun-
try Trends, Methods and Statistics. Geneva: acc/scn in collaboration with the Inter-
national Food Policy Research Institute.
------. 2000. Fourth Report on the World Nutrition Situation. Geneva: acc/scn in col-
laboration with the International Food Policy Research Institute.
Alderman, Harold. 2000. "Anthropometrics." In Margaret Grosh and Paul Glewwe, eds.,
Designing Household Survey Questionnaires for Developing Countries: Lessons from
Ten Years of LSMS Experience. New York: Oxford University Press.
Alderman, Harold, and Marito Garcia. 1994. "Food Security and Health Security: Ex-
plaining the Levels of Nutritional Status in Pakistan." Economic Development and
Cultural Change 42(3):485�508.
Allen, Linsey, and Stuart Gillespie. 2001. What Works? A Review of the Efficacy and
Effectiveness of Nutrition Interventions. adb Nutrition and Development Series, no.
5. Geneva: acc/scn, and Manila: Asian Development Bank.
Anand, Sudhir, and Martin Ravallion. 1993. "Human Development in Poor Countries:
On the Role of Private Income and Public Services." Journal of Economic Perspec-
tives 7(winter):133�50.
Berg, Alan. 1981. Malnourished People: A Policy View. Washington, D.C.: World Bank.
Bouis, Howarth, and Lawrence Haddad. 1992. "Are Calorie-Income Elasticities Too
High? A Recalibration of the Plausible Range." Journal of Development Economics
39:333�64.
Bound, John, David A. Jaeger, and Regina M. Baker. 1995. "Problems with Instru-
mental Variables Estimation When the Correlation between the Instruments and the
Endogenous Explanatory Variable Is Weak." Journal of the American Statistical
Association 90(430):443�50.
Burger, Susan, and Steven Esrey. 1995. "Water and Sanitation: Health and Nutrition
Benefits to Children." In Per Pinstrup-Andersen, David Pelletier, and Harold Alder-
man, eds., Enhancing Child Growth and Nutrition in Developing Countries: Priori-
ties for Action. Ithaca, N.Y.: Cornell University Press.
Davidson, Russell, and James G. MacKinnon. 1993. Estimation and Inference in Econo-
metrics. New York: Oxford University Press.
Deaton, Angus. 2001. "Counting the World Poor: Problems and Possible Solutions."
World Bank Research Observer 16(2):125�48.
Gillespie, Stuart, and Lawrence Haddad. 2001. Attacking the Double Burden of Malnu-
trition in Asia and the Pacific. adb Nutrition and Development Series, no. 4. Geneva:
acc/scn, and Manila: Asian Development Bank.
Gillespie, Stuart, John Mason, and Raynaldo Martorell. 1996. How Nutrition Improves.
Geneva: acc/scn.
Greene, W. H. 2000. Econometric Analysis, 4th ed. Upper Saddle River, N.J.: Prentice Hall.
Griliches, Zvi, and Jerry Hausman. 1986. "Errors in Variables in Panel Data." Journal
of Econometrics 31:93�118.
Grosh, Margaret, and Paul Glewwe, eds. 2000. Designing Household Survey Question-
naires for Developing Countries: Lessons from Ten Years of LSMS Experience. New
York: Oxford University Press.
Haddad and others 131
Harriss, Barbara. 1995. "The Intrafamily Distribution of Hunger in South Asia." In Jean
Dr�ze, Amartya Sen, and Artar Hussain, eds., The Political Economy of Hunger: Se-
lected Essays. Oxford: Clarendon.
Johnston, John, and John DiNardo. 1997. Econometric Methods, 4th ed. New York:
McGraw-Hill.
Krueger, Alan, and Mikael Lindhal. 1999. Education for Growth: Why and for Whom.
Industrial Relations Section Working Paper 429. Princeton University, Princeton, N.J.
Maxwell, Simon. 1999. "International Targets for Poverty Reduction and Food Security:
A Mildly Skeptical but Resolutely Pragmatic View with Greater Calls for Subsidiarity."
IDS Bulletin 30(2):92�105.
Pelletier, David. 1994. "The Potentiating Effects of Malnutrition in Child Mortality:
Epidemiologic Evidence and Policy Implications." Nutrition Reviews 52(12):409�15.
Pritchett, Lant, and Lawrence Summers. 1996. "Wealthier Is Healthier." Journal of
Human Resources 31(4):841�68.
Reutlinger, Shlomo, and Marcelo Selowsky. 1976. Malnutrition and Poverty: Magnitude
and Policy Options. Baltimore, Md.: Johns Hopkins University Press.
Sahn, David, David Stiffel, and Stephen Younger. 1999. Intertemporal Changes in Wel-
fare: Preliminary Results from Nine African Countries. Cornell Food and Nutrition
Policy Program Working Paper 94. Ithaca, N.Y.: Cornell University.
Smith, Lisa, and Lawrence Haddad. 2000. Explaining Child Malnutrition in Develop-
ing Countries: A Cross-Country Analysis. Research Report 111. International Food
Policy Research Institute, Washington, D.C.
------. 2002. "How Potent Is Economic Growth in Reducing Undernutrition? What Are
the Pathways of Impact? New Cross-Country Evidence." Draft. International Food
Policy Research Institute, Washington, D.C.
Strauss, John, and Duncan Thomas. 1995. "Human Resources: Empirical Modeling of
Household and Family Decisions." In Jere R. Behrman and T. N. Srinivasan, eds.,
Handbook of Development Economics, vol. 3A. Amsterdam: North Holland.
------. 1998. "Health, Nutrition and Economic Development." Journal of Economic
Literature 36:766�817.
undp (United Nations Development Programme). 2001. Human Development Report
2001. New York: Oxford University Press.
who (World Health Organization). 1995. Physical Status: The Use and Interpretation
of Anthropometry. Technical Report Series 854. Geneva.
------. 1997. Global Database on Child Growth and Malnutrition. Geneva.
World Bank. 1997. World Development Indicators 1997. Washington, D.C.
------. 1998. World Development Indicators 1998. Washington, D.C.
------. 2002. Deininger and Squire Data Set. Available online at www.worldbank.org/
research/growth/dddeisqu.htm.
the world bank economic review, vol. 17, no. 1 133�143
Particularism around the World
Jessica Seddon Wallack, Alejandro Gaviria,
Ugo Panizza, and Ernesto Stein
This article presents a new data set on electoral systems and outlines its potential
uses in research on the links between electoral systems and economic outcomes. The
data measure the extent to which politicians can advance their careers by appealing
to narrow geographic constituencies on the one hand or party constituencies on the
other.
Electoral systems have long been viewed as mechanisms that enforce politicians'
accountability. There has been less systematic focus, however, on differences in
the entities to whom these elected representatives are accountable. The data set
outlined in this article begins to quantify the varying incentives that electoral
systems around the world create. In particular, the variables in the data set indi-
cate the extent to which the electoral process creates incentives for politicians to
cater to narrow constituencies.
The data set is useful for several areas of political economy. Differences in the
effective arbiters of policymakers' careers may influence how different interest
groups can affect policymaking. In systems where politicians' careers are deter-
mined by the wills (and whims) of their constituencies, interest groups must
channel their demands through district-level politics. In systems where candi-
dates' futures are determined by party favors, interest groups may gain more
influence by appealing to the party leaders who oversee politicians.
The ways in which politicians further their careers are also likely to influence
their policymaking priorities. For example, the strength of their connections to
Jessica Seddon Wallack is a graduate student in political economics in the Graduate School of Busi-
ness at Stanford University; her e-mail address is jseddon@stanford.edu. Alejandro Gaviria is Deputy
Director of the Planning Department, Colombia at Fedesarollo; his e-mail address is agaviria@dnp.gov.co.
Ugo Panizza is Economist in the Research Department at the Inter-American Development Bank; his
e-mail address is ugop@iadb.org. Ernesto Stein is Principal Economist in the Research Department at
the Inter-American Development Bank; his e-mail address is ernestos@iadb.org. The authors are grate-
ful for useful comments from Fran�ois Bourguignon, Jeffry Frieden, Stephen Haggard, Matthew Shugart,
and participants in seminars at Stanford University, the Inter-American Development Bank, Harvard
University's Center for Basic Research in the Social Sciences, and the April 6�8, 2000, Latin American
and Caribbean Economic Association�Political Economy Group meeting in Cartagena, Colombia. The
views expressed in this article do not necessarily reflect those of Fedesarollo or the Inter-American
Development Bank.
DOI: 10.1093/wber/lhg010
� 2003 The International Bank for Reconstruction and Development / THE WORLD BANK
133
134 the world bank economic review, vol. 17, no. 1
electoral districts can have important policy consequences.1 In addition, differ-
ences in politicians' incentives to appeal to narrow geographic constituencies or
to party policies may explain cross-country differences in the division of public
spending between broad and targeted programs (Lizzeri and Persico 2001).
Milesi-Ferreti and others (2002), for example, find that party-oriented govern-
ments tend to have higher transfers and lower shares of public goods in govern-
ment spending (see also Persson and Tabellini 2002).
Finally, the incentives that are the focus of this article affect policymaking in
legislatures. Systems where each politician is attentive to narrow interests are
likely to foster gridlock in legislatures. At the same time, party-centered systems
may lack the institutional channels for competing views to be expressed and
resolved in sustainable policies. Particularistic systems build incentives for legis-
lators to gather information on the preferences of their constituencies and may
also generate competition among legislators to serve constituents better. The
importance of constituency approval, on the other hand, might motivate legis-
lators to claim credit and so complicate policymaking. Several studies have found
that the middle ranges of particularistic incentives--countries in which legisla-
tors must balance the demands of their constituents and their parties--are linked
to better policy outcomes, including faster recovery from crisis (Gaviria and others
2000), easier economic reform (Shugart 2001), and higher-quality institutions
(Panizza 2001).
The data set presented in this article is a useful complement to existing data
on political institutions. It expands on the measures of incentives for particular-
ism provided in the Database of Political Institutions compiled by Beck and oth-
ers (2001). Although that database includes broad characteristics of electoral
systems, this article draws on theoretical work by Carey and Shugart (1995) to
add more specific indicators of the strength of incentives for attention to nar-
row constituencies. In addition, detailed separate data are provided for rules that
affect the incentives of candidates for upper and lower chambers of legislatures,
a level of disaggregation useful for researchers interested in policy areas in which
one house dominates the other.
The data described in this article can also be used to further differentiate
among democracies as identified in broader data sets such as the Polity IV data
set.2 Though Polity IV includes indicators on the general competitiveness of
political systems and the process for selecting executives, the data presented
here provide a far more nuanced view of the cross-country variations in the
incentives facing legislators.
1. The data are best suited for measuring incentives to cater to geographically concentrated special
interests, because these are most likely to coincide with electoral districts. The current data set does not
consider the equally interesting question of the strength of incentives to curry favor with narrow inter-
est groups that are geographically dispersed.
2. The data set can be obtained online at www.cidcm.umd.edu/inscr/polity/index.htm#data.
Wallack and others 135
I. The Data Set
The data set has a panel structure and covers a maximum of 158 countries from
1978 to 2001. The panel structure of the data can be useful for analysts of elec-
toral reform, institutional change, and comparative institutions. Whereas much
of the literature on the determinants of electoral change is drawn from individual
country experiences, this data set provides a dependent variable for testing some
of these theories across countries.
The main sources of data are the Inter-Parliamentary Union (ipu) Chronicle
of Parliamentary Elections and Developments (various years) and online Parline
database.3 The International Institute for Democracy and Electoral Assistance
(idea) Handbook of Electoral System Design (1997) and the Parlamento Latino-
americano Manual de los partidos pol�ticos de America Latina (1997) were also
consulted when ipu information was incomplete. District magnitude data from
the ipu were supplemented by table 3.2 in Cox (1997).
The data set covers countries with varying degrees of civil liberties and politi-
cal rights. Needless to say, electoral systems are less relevant for policy outcomes
in dictatorships than in democratic regimes (for instance, one could argue that it
does not make much sense to measure electoral incentives in Latin America or
Sub-Saharan Africa in the 1980s). Furthermore, even in formally democratic
countries, corruption, interest group pressures, and other factors may--in tan-
dem with the need to get votes--influence legislators' behavior. Electoral pres-
sures are a subset, though an important one, of the incentives facing legislators.
Readers can define the level of autocracy at which legislators' incentives become
irrelevant for policy outcomes.4 The data set includes an indicator variable for
one-party states.
The data set can be downloaded in Stata and Excel formats from www.stanford
.edu/~jseddon. The next section discusses the rationale behind the coding of the
variables, and the Excel workbook includes more detailed notes on classifica-
tions of more complicated electoral systems. Appendix tables A-1 and A-2 list
the files available on the Internet and the variables in the data set.
II. Description of Variables
The collection of the data on electoral formulas was guided by Carey and
Shugart's (1995) theoretical work on the incentives that different electoral for-
mulas create to cultivate a personal vote, as well as by Shugart's (2001) work
on the links between economic and electoral reforms. The four key character-
3. Current information on electoral systems is available online at www.ipu.org/parline-e/
parlinesearch.asp.
4. Data on democracy are available from the Polity IV data set, available online at www.bsos.umd.edu/
cidcm/polity/#data.
136 the world bank economic review, vol. 17, no. 1
istics of electoral systems proposed by Carey and Shugart and recorded (with
some modifications to their framework) in this data set are:
� Ballot--party and citizen control over candidates' access to and location
on ballots;
� Pool--extent to which candidates can draw on their parties' reputations
to win elections;
� Vote--number and specificity of votes; and
� District magnitude.5
The various dimensions of particularism are positively correlated but not iden-
tical: the correlation between the Ballot and Pool indicators is 0.60, between
Ballot and Vote 0.74, and between Pool and Vote 0.63.
As in Carey and Shugart, the electoral system indicators described here range
from zero (for systems where politicians' careers depend most on party fortunes)
to two (for the most particularistic systems, where candidates must focus on
narrow geographic constituencies).6 In addition, the data set includes the pro-
portion of legislators from a national constituency (PropN). As noted, separate
indicators are reported for upper and lower houses in bicameral systems, but
researchers can easily create a composite country value of the variables by aver-
aging the values of the two houses or assigning other weights based on the houses'
importance in policymaking.
Ballot describes the relative strengths of parties and citizens in shaping candi-
dates' access to ballots and influencing their chances of being elected.7 Electoral
systems in which parties control candidates' positions on ballots give parties the
most control over entry into politics. These systems, generally known as closed
list electoral systems, are coded as zero. Politicians in these systems have a strong
incentive to cater to parties rather than constituents to be chosen and placed in
a viable spot near the top of the list.
Systems are coded as one if parties exert strong influence over which candi-
dates are on the ballot but do not control the order in which candidates ap-
pear. Open list systems, in which voters can rank candidates on a party-selected
list, are in this category. Systems in which independent candidates are legal
but there are high formal or informal barriers to getting one's name on the
ballot are also in this category. Politicians in this group must balance efforts
5. No version of Carey and Shugart's index of particularism is reported here because that index is
simply a summation of Ballot, Pool, and Vote. It is not clear that summation is an appropriate way to
capture the degree of particularism in a country, and this index is at best an ordinal measure of particu-
larism. Weights derived from principal component analysis might be more appealing for researchers
seeking a summary statistic for a country, but this article avoids that kind of distillation of what is
already a reduction of complex laws to a few simple measures.
6. It is important to remember that the implications for policy outcomes of a "party-centered sys-
tem" are difficult to discern without further knowledge about how parties affect policymaking.
7. Although Ballot is similar to the variable CL in Beck and others (2001), a finer classification is
used here.
Wallack and others 137
to please the party with efforts to attract constituent support. Systems with
low independent candidacy requirements and plurality thresholds (as opposed
to a stricter absolute majority requirement) offer voters more influence over
the selection of candidates and are coded as two. Candidates in this group focus
exclusively on gaining support among their constituents, and there is little need
to gain party favor.
Both formal and informal entry barriers were considered when distinguish-
ing between codes of one and two. The assessment of formal barriers is based
on legal restrictions, such as mandatory party membership, whereas the assess-
ment of informal barriers is based on the history of successful independent can-
didates in supplementing scarce information on party nomination procedures.8
Political entrepreneurs will have little incentive to adhere to party rules if they
can easily bypass them.
The main departure from the classification scheme proposed by Carey and
Shugart (1995) was in coding single-member districts. Carey and Shugart see all
single-member districts as closed-party lists of one for smaller districts and code
them all as zero. We differentiate between single-member districts based on their
context to avoid overemphasizing the role of parties (as opposed to voters) in
selecting candidates. Candidates' popularity with voters in such districts is likely
to be more important in gaining access to a list of one than a list of several. Thus,
single-member districts are coded as zero in countries for which the majority of
other districts are multimember closed-list proportional or where there was a
single-party system (as in Bulgaria from 1981�89, Mali, or Sierra Leone). Other
single-member districts were assigned a value of one, indicating that parties re-
tain some control over ballots but voters can influence party choices in coun-
tries where closed lists do not predominate. The data set includes a dummy
variable for single-party legislatures.
Pool measures the extent to which candidates can ride their parties' reputa-
tions to electoral success. In systems where votes are pooled across candidates,
the electoral success of the party determines individuals' careers. Candidates thus
have little incentive to build personal bases of support. Candidates who receive
no spillover votes from party colleagues, on the other hand, will compete harder
to create personal support bases. Here this variable is seen from the candidate's
perspective, with consideration given to whether candidates for national office
can expect to benefit from electoral support for other candidates in their party,
possibly in other districts.
8. The history of independent candidatures is an imperfect indicator of entry barriers because it is
an outcome variable and is not based on the same kinds of institutional data as other variables. But
casual knowledge of electoral politics suggests that relatively unobservable factors (such as campaign
financing, social pressure, and restrictions on advertising) can impede independent candidates as much
as (if not more than) formal requirements for running for office. The United States, for example, has
relatively low formal barriers to independent candidacy, but the costs of running a campaign and re-
strictions on fundraising make party affiliation a near necessity. It is important to recognize these kinds
of barriers in some way. Cases where outcome variables have been used are highlighted.
138 the world bank economic review, vol. 17, no. 1
Systems that pool votes across all candidates in a party are coded as zero. Can-
didates in these settings attain seats in the legislature if their party attracts votes,
regardless of the level of personal support they attract from voters. Systems where
votes are pooled across a subset of the party are coded as one. The group among
which votes are shared is smaller, increasing the reward for attracting personal
support. Electoral systems where voters can direct support to individual candi-
dates are coded as two. Candidates have a greater incentive to attract personal
support because their colleagues' popularity will not earn them any votes.
The coding of the Pool variable used here diverges substantially from that in
Carey and Shugart (1995). Carey and Shugart define Pool according to whether
votes for a candidate contribute to the probability of others in his or her party
winning seats in that electoral district. This difference in coding is most obvious
in the case of single-member districts. Carey and Shugart classify most single-
member districts as having a Pool code equal to zero because each candidate is
presented as a list of one, and votes for the candidate are thus pooled across the
entire list. Our definition, in contrast, causes candidates in single-member dis-
tricts to receive codes of two on the Pool scale because they do not receive addi-
tional electoral support if other candidates from their party are successful in other
districts. We feel that the code of two more accurately reflects the incentives facing
candidates in single-member districts. Again, single-party states where candidates
often stand for election as local representatives of a party are an exception to
this rule and are coded as zero.
Vote measures limitations on the number of candidates that voters can sup-
port. Legislators have a stronger incentive to please their constituencies if the
number of votes is limited and they must convince voters to choose only them.
They will have little incentive to cater to their home constituency if they cannot
attract votes individually but only as a party member.
As in Carey and Shugart (1995), the values range from zero for a single vote
for a party to one for multiple votes across candidates (who may or may not be
from the same party) to two for a single vote for a single candidate. Electoral
systems where voters cast two votes--one for a local candidate and another for
a national candidate--are coded as one. Multiple votes may also be spread over
time, as in systems where there are multiple rounds of elections to narrow the
field of candidates. While systems with open primary elections are counted as
having multiple votes, systems with rare tie-breaking runoffs are not. Candidates
in the latter systems do not regularly expect to have to expand their audiences
after the first round. Single-member districts are still coded as two (again in
contrast to Carey and Shugart) except in single-party states, because people are
voting for a candidate. This coding is consistent with the method used to code
Ballot (see above).
District magnitude may also affect how legislators build their personal repu-
tations. Larger districts are likely to increase the need for legislators to internal-
ize the consequences of redistributive policies. It is harder to find policies that
do not create both losers and winners in larger districts, and particularistic dis-
Wallack and others 139
tributive policies targeted to a narrow constituency are less likely to be successful
in attracting votes (Lancaster 1986). But larger districts also increase the inten-
sity of competition between candidates and the need for candidates to differen-
tiate themselves from others (Cox 1990). Because it is difficult to determine a
priori which effect dominates, this variable is included without predictions about
its effect on policymaking or policy outcomes.9
As noted, the coding of single-member districts is perhaps the most significant
departure from the framework outlined in Carey and Shugart (1995) for the vari-
ables Ballot, Pool, and Vote. An indicator variable is included to show the pro-
portion of single-member districts (smd), and such districts are noted in the
background file so that users who prefer to interpret single-member districts as
closed lists of one can change the coding accordingly.
Finally, a variable is included for the proportion of legislators from national
constituencies in each house (PropN). This variable can be seen as a broad sum-
mary measure of incentives to cater to narrow constituencies. PropN summa-
rizes incentives for all legislators in a country, both elected and appointed, because
it is based only on the identity of who selects each candidate rather than the
process used to select them. Candidates appointed by national leaders are con-
sidered to be from national constituencies, and candidates indirectly elected by
provincial legislators or appointed by subnational councils are considered to have
nonnational constituencies. These politicians' loyalties to a geographically de-
fined support base are not clear.
III. Descriptive Statistics
This section presents some brief descriptive statistics of the averages of the main
variables over time (table 1) and of the broadest measure, PropN, by region (table
2). Although the values for individual countries can change markedly with elec-
toral reforms, the averages of the variables for the lower or only house and the
upper house (H2) are fairly stable over time. There is little evidence of an overall
trend toward more or less personalistic incentives in electoral systems. The num-
ber of observations changes over time, particularly in the 1980s, as more coun-
tries move to at least nominal use of electoral systems.
One interesting point from table 1 is that upper houses (H2) tend to be more
party-centered than lower or only houses. The averages for Ballot and Pool tend
to be slightly lower, indicating more party control over access to ballots and
stronger incentives to free ride on a party's reputation rather than seek personal
support. But the most marked difference between upper and lower or only houses
is that upper houses tend to have a larger proportion of representatives from
national constituencies.
9. Carey and Shugart (1995) hypothesize that the incentive to cultivate a personal reputation in-
creases with district magnitude in candidate-centered systems and decreases with district magnitude in
closed-list systems, in which parties determine who is on the ballot and what position they are in.
Table 1. Averages of the Main Variables, 1978�2001
No. of District
No. of observations Ballot Pool Vote District magnitude PropN
Year observationsa (H2) Ballot (H2) Pool (H2) Vote (H2) magnitude (H2) PropN (H2)
1978 59 16 0.77 0.69 0.92 1.00 1.06 1.13 12.95 6.10 0.13 0.34
1979 75 16 0.78 0.69 1.01 1.00 1.06 1.13 15.11 12.93 0.14 0.39
1980 88 18 0.75 0.67 1.03 0.96 1.01 1.06 17.62 16.07 0.16 0.40
1981 92 18 0.77 0.67 1.09 0.96 1.02 1.06 17.30 16.78 0.16 0.42
1982 91 18 0.77 0.67 1.08 0.96 1.02 1.06 14.73 16.78 0.15 0.42
1983 101 21 0.74 0.57 1.03 1.01 0.99 1.10 14.08 14.38 0.14 0.39
1984 112 24 0.72 0.63 1.03 1.09 1.00 1.13 13.41 13.78 0.14 0.42
1985 113 24 0.78 0.71 1.07 1.01 1.07 1.08 13.37 14.19 0.14 0.42
1986 113 24 0.78 0.71 1.07 1.01 1.07 1.13 12.56 14.63 0.13 0.44
1987 120 24 0.78 0.75 1.07 1.01 1.09 1.13 13.60 14.63 0.13 0.43
140 1988 117 22 0.77 0.65 1.08 0.92 1.09 1.09 13.87 15.07 0.12 0.45
1989 124 24 0.76 0.62 1.07 0.93 1.05 1.04 13.06 13.22 0.12 0.43
1990 122 25 0.78 0.63 1.06 0.89 1.11 1.08 11.98 12.96 0.11 0.39
1991 118 27 0.77 0.66 1.05 0.90 1.10 1.11 11.22 11.79 0.11 0.38
1992 131 29 0.79 0.71 1.05 0.97 1.13 1.24 10.93 11.05 0.10 0.38
1993 143 30 0.81 0.71 1.06 0.94 1.15 1.20 10.08 11.33 0.10 0.40
1994 146 29 0.79 0.69 1.04 0.88 1.14 1.08 10.22 12.42 0.09 0.41
1995 153 29 0.80 0.69 1.05 0.88 1.13 1.08 11.26 12.64 0.11 0.42
1996 156 28 0.79 0.68 1.07 0.84 1.16 1.09 11.33 12.70 0.11 0.42
1997 158 29 0.78 0.69 1.05 0.95 1.14 1.12 11.57 12.43 0.12 0.40
1998 154 29 0.78 0.69 1.05 0.95 1.13 1.12 11.86 11.73 0.12 0.42
1999 155 30 0.78 0.70 1.05 0.99 1.15 1.15 11.90 11.49 0.12 0.42
2000 155 30 0.78 0.70 1.05 1.12 1.15 1.21 11.65 11.60 0.12 0.43
2001 154 30 0.78 0.70 1.06 1.12 1.16 1.21 11.77 11.60 0.12 0.41
Note: Unspecified data cover lower houses of the legislature or countries with only one house. Data marked with H2 cover upper houses of the legislature.
aThe number of observations for Ballot, Pool, and Vote occasionally vary for a given year because it was not always possible to find all the information
needed for each variable. The data in this column are the number of countries with observations for all three variables in each year.
Source: Authors' calculations.
Wallack and others 141
Table 2. Average Proportion of Legislators (PropN)
from National Constituencies by Region, 1980, 1990,
and 2000
Region 1980 1990 2000
High income 0.143 0.140 0.168
Latin America and Caribbean 0.227 0.241 0.254
Sub-Saharan Africa 0.223 0.214 0.178
East Asia 0.361 0.121 0.132
South Asia 0.013 0.052 0.079
Middle East and North Africa 0.200 0.192 0.118
Eastern Europe and Central Asia -- 0.041 0.220
Note: Data are for countries with bicameral legislatures. Regions and
income groups are defined using World Bank classifications. Countries clas-
sified as high income are not included in their respective regions.
Source: Authors' calculations.
In table 2, one caveat to keep in mind is that having a national constituency
need not mean that a candidate was elected; he or she may have been appointed
by a national leader.10 Latin American and Sub-Saharan Africa stand out as
having the highest proportions of legislators from national constituencies,
whereas South Asian countries tend to have the most representatives from smaller
constituencies. Eastern Europe and Central Asia's average proportion of legisla-
tors from national constituencies has increased as many transition economies
have adopted proportional electoral systems with at least some seats from na-
tional pools. The average proportion from national constituencies has also in-
creased among high-income countries. This trend appears to be the result of the
increasingly common practice of allocating seats after elections to ensure a dis-
tribution of seats that is more proportional to the number of votes that parties
have received.
It is important to remember, however, that the impact of these electoral incen-
tives relative to corruption, interest group pressures, and other nonelectoral de-
terminants of politicians' careers vary across countries. Although these descriptive
statistics highlight rough differences across regions and some trends over time, the
overall averages mask nontrivial changes in countries' electoral systems and im-
portant differences across countries with similar scores on individual variables. It
is important to consider the country-specific information in the data set documen-
tation as well as the various control variables mentioned in section II.
IV. Conclusion
There is growing consensus that political institutions play an important role in
shaping a country's economic policies. Empirical work on this subject, however,
10. The country notes accompanying the data in the Excel spreadsheet and the indicator variable
for proportion of directly elected legislators should help researchers distinguish between these cases.
142 the world bank economic review, vol. 17, no. 1
has been hampered by a lack of detailed data for a large set of countries over
time. This article helps narrow that gap by operationalizing an intuitively ap-
pealing theoretical framework for measuring legislators' incentives. It is a comple-
ment to larger data sets such as Polity IV and the Database of Political Institutions
developed by Beck and others (2001), because it provides a more nuanced way
of differentiating between democracies. Although the data set described in this
article is far from being a complete inventory of how electoral systems affect
political incentives, it may be useful in providing the tools to test some of the
relationships between institutions and economic outcomes that have been high-
lighted in the theoretical literature.
Appendix
Table A-1. Files Available in the Data Set
File name Format and content
Public2001.dta State data set with full panel of data
Public2001.xls Excel workbook
Electoral Data Spreadsheet of coded values
Country Notes Notes on country-specific ambiguities in coding
Coding Notes on general coding decisions
Source: www.stanford.edu/~jseddon.
Table A-2. Names and Descriptions of Variables in the Data Set
Variable name Description
COUNTRY Country name
SHCODE Country code
BICAMERAL Dummy variable, 1 if bicameral system
YEAR Year
ONEPARTY Dummy variable, 1 if single-party system
BALLOT Party control over access to and position on ballot, lower/only house
BALLOT2 Party control over access to and position on ballot, upper house
POOL Sharing of votes among candidates of the same party, lower/only house
POOL2 Sharing of votes among candidates of the same party, upper house
VOTE Candidate- or party-specific voting, lower/only house
VOTE2 Candidate- or party-specific voting, upper house
CINDEX Proportion of legislators included in the index, lower/only house
CINDEX2 Proportion of legislators included in the index, upper house
DM District magnitude, lower/only house
DM2 District magnitude, upper house
SMD Proportion of legislators from single-member districts, lower/only house
SMD2 Proportion of legislators from single-member districts, upper house
PROPN Proportion of legislators from national constituencies, lower/only house
PROPN2 Proportion of legislators from national constituencies, upper house
Source: www.stanford.edu/~jseddon.
Wallack and others 143
References
Beck, Thorsten, George Clark, Alberto Groff, Philip Keefer, and Patrick Walsh. 2001.
"New Tools in Comparative Political Economy: The Database of Political Institutions."
World Bank Economic Review 15:165�76.
Carey, John, and Matthew Soberg Shugart. 1995. "Incentives to Cultivate a Personal Vote:
A Rank Ordering of Electoral Formulas." Electoral Studies 14(4):417�39.
Cox, Gary. 1990. "Centripetal and Centrifugal Forces in Electoral Systems." American
Journal of Political Science 34(4):903�35.
------. 1997. Making Votes Count: Strategic Coordination in the World's Electoral
Systems. Cambridge: Cambridge University Press.
Gaviria, Alejandro, Ugo Panizza, Jessica Seddon, and Ernesto Stein. 2000. "Political
Institutions and Growth Collapses." Working Paper 419. Inter-American Develop-
ment Bank, Research Department, Washington, D.C.
International idea (Institute for Democracy and Electoral Assistance). 1997. Handbook
of Electoral System Design. Stockholm: International idea.
ipu (International Parliamentary Union). Various years. Chronicle of Parliamentary Elec-
tions and Developments. Geneva.
Lancaster, Thomas. 1986. "Electoral Structures and Pork Barrel Politics." International
Political Science Review 7(1):67�81.
Lizzeri, Alessandro, and Nicola Persico. 2001. "The Provision of Public Goods under
Alternative Electoral Incentives." American Economic Review 91(1):225�45.
Milesi-Ferreti, Gian-Maria, Roberto Perotti, and Massimo Rostagno. 2002. "Electoral
Systems and Public Spending." Quarterly Journal of Economics 117(2):609�57.
Panizza, Ugo. 2001. "Electoral Rules, Political Systems, and Institutional Quality." Eco-
nomics and Politics 13:311�42.
Parlamento Latinamericano. 1997. Manual de los Partidos Politicos de America Latina.
S�o Paulo.
Persson, Torsten, and Guido Tabellini. 2002. "Political Institutions and Policy Outcomes:
What Are the Stylized Facts?" Working paper, Innocenzo Gasparini Institute for Eco-
nomic Research, Bocconi Univerisity, Milan, Italy.
Shugart, Matthew. 2001. "Extreme Electoral Systems and the Appeal of the Mixed-
Member Alternative." In Matthew Shugart and Martin Wattenburg, eds., Mixed-
Member Electoral Systems: The Best of Both Worlds? New York: Oxford University
Press.
American Law and Journalof Economic
Econom~csRev~ew Geography
Cambridge Journai Journal of International
of Economics Economic Law
The Journalof Law,
Contemporary
Economics, &
Econom~cPolicy
Organization
Contributions to
Oxford Economic Papers
Political Economy
Oxford Review of
Economic Inquiry Economic Policy
Enterprise & Soci~ty: The Review of Financial
The Initernat~onal Studies
journal of BUS~IIESS
Social Science Japan
W 1story
Journal
European Review
Socio-Economic Review
of Agricultural
.: Econonlics The World Bank
Economic Review
: Industrial and
Corporate Change The World Bank
Research Observer
$ Journal of African
Econon:ies
&f?{
~$2
%Y3" Visit our Web site today for more
Informaition or to register to receive the
table of contlents FREE by e-mail!
Coming in the next issue of
THE WORLD BANK
ECONOMIC REVIEW
;Vblume 17,Number 2,2003
CHILDLABOR AND DEVELOPMENT
Introduction
Kaushik Basu and Zafiris Tzannatos
The Global Chid Labor Problem: What do we know and
what canwe do?
Kaushik Basu and Zafiris Tzannatos
Chilld Labor: The Experience of Today's Advanced Economies
and Lessons from the Past
Jane Humphries
Child Farm Labor: The Wealth Paradox
Sonia Bhalotra and ChristopherHeady
Conditional Cash Transfers, Schooling, and Child Labor:
Micro-Simulating Bolsa Escola
Frangois Bourguignon, Francisco H.G. Ferreira, and Phil(ippe G. Leite
Targeting Child Labor in Debt Bondage: Evidence,
Theory, and Policy Implications
Arnab K. Basu and Nancy H. Chau
Children's Worlking Hours and SchoolEnrollment: Evidence
from Pakistan and Nicaragua
Furio CamilloRosati and Mariacristina Rossi
Child Labor: A Normative Perspective
Debra Satz