77403 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 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 40.7 110 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 t-statistic 3.13 3.10 2.11 1.51 Hausman testb p = 0.027 p = 0.393 114 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 t-statistic 3.28 2.00 2.78 2.89 Hausman testb p = 0.279 p = 0.066 115 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 Percentage of children under five who are 30 underweight 25 1970's 20 1980's 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 enterprise, and agricultural machinery (tractors, threshers) (and × rural-urban dummy variable) 126 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 white TV Mozambique 127 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 Peru Estimated coefficienta 0.2504 1.2001 0.2056 0.8150 128 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) Hausman test, ols vs. ivb p = 0.000 p = 0.000 p = 0.057 p = 0.049 129 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 Drèze, 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.