THE WORLD BANK ECONOMIC REVIERWl Volume 10 January 1996 Number 1 Hoxv Important to India's Poor Is the Sectoral C(mposition of Economic Growth? .\'lartin RavlIlion and Gaurav Datt Is the Debt Crisis History? Recent Private Capital Inflows to Developing Countries Michael Dooley, Eduardo Fernandez-Arias, and Kenneth Kletzer The Surge in Capital Inflows to Developing Countries: An Analytical Overview Eduardo Fernindez-Arias and Peter J. Montiel A SYNtri(oSIU\I O(N FERTILITY IN SUB-SAHARAN AFRICA Introduction: Fertility in Sub-Saharan Africa Martha Ainsworth The Impact of Women's Schooling on Fertility and Contraceptive Use: A Study of Fourteen Sub-Saharan African Countries r Martha Ainswortih, Kathleen Beegle, and Andrew Nyamere Fertility and Child Mortality in C6te dilvoire and Ghana Kofi Benefo and T. Paul Schultz Contraceptive Use and the Quality, Price, and Availability of Family Planning in Nigeria Bamilkale J. Fevisetan and lMlartha Ainsworth Fertility, Contraceptive Choice, and Public Policy in Zimbabwe Duncan Thomas and John Mvaluccio THE WORLD BANK ECONOMIC REVIEW EDITOR Nloshe Syrquin CONSULi_ 1N, (,ID1] oR Sandra GaIn EDIIORIAL BOARD Kaushik Basu, Cornell University and University of Delhi David Dollar Guillermo Calvo, University of Maryland Gregory K. Ingram Jonathan Eaton. Boston University John Page Alberto Giovannini, Rome. Italy Lant H. Pritchett Mlark R. Rosenzweig, University of Pennsylvania Jacques van der Gaag Th7e World Batik Econo)mic Revietv is a professional journal for the dissemination of World Bank- sponsored research that informs policv analvses and 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 emphasizes policy relevance and opera- tional aspects of economics, rather than primarilv 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 policv will nor be grounds for rejection of an article. Articles will be drawn primarily from work conducted by World Bank staff anid consultants. Before being accepted for publication by the Editorial Board, all articles are reviewed hy two referees who are not mem- bers of the Bank's sraff and one World Bank staff member; articles niust also be recommended bv at least one external member of the Editorial Board. The Ret'iew may on occasion publish articles on specified topics by non-Bank contributors. Any reader interested in preparing such an article is invited to submit a proposal of not more than two pages in length to the Editor. The views and interpretations expressed in this journial are those of the authors and do not necessarily represent the views and policies of the World Bank or of its Executive Directors or the countries they rep- resent. The World Bank does not guarantee the accuracv of the data included in this publication and accepts no responsibility whatsoever for any conisequiences of their use. When maps are used, the bound- aries, denominations, and other information do nor imply on the part of the World Bank Group any judg- ment on the legal status of any territory or the endorsement or acceptance of such boundaries. Comments or brief notes responding to Rcviewt, articles are xvelcome and will be considered for publica- tion to the extent that space permits. Please direct all editorial correspondence to the Editor. Tlhe World Ba.nk Economic Rcv'ieie, The World Bank, WVashinigton, D.C. 10433, U.S.A. 117e World Bank Economic Reziew' is published rhree rimes a year (January, May, and September) by rhe World Bank. Single copies may be purchased at $1 0.95. Subscription rates are as follows: l,rdoil dlals Institoetionis 1-year subscription UlS$25 US$45 2-vear subscription USS46 US$86 i-yearsubscription US$65 US$125 Orders should be sent to: World Bank PLublications, Box -247-7956, Philadelphia, PA 19170-7956, U.S.A. Subscriptions are available without charge to readers wNith mailingaddresses in developing countries and in socialist economies in transition. Written request is required every two years to renew such subscriptions. ©D 1996 The international Bank for Reconstruction and Development / Tli I WORL ) BANK 1818 H Street, N.W.. Washington, D.C. 20433, U.S.A. All rights reserved M.aJilfacrLred in the United States of America ISSN (025S-677l Mlaterial in this jourinal is copyrighted. Requesrs to reproduce portionis of it should be sent to the Office of the Publisher at the address in the copyright notice above. The World Bank encourages dissemination of its wvork and wvill normally give permissioin promprlv and. ss hen the reproduction is for iioiiconsilercial purposes, withouit asking a fee. Permission to maike photocopies is granted through the Copyright Clearance Center, Suite 91 0, 222 Rosewood Drive, Danvers, M0A 019213, l.S.A. This journal is indexed regularly in Corrre rt Contemtts/Social & Behavioral Sciernces, Ildex to Inrterna- tional Statis tics. Ion rnal of Economic L itera t ure, Public A4falirs I ntorma tion Sen ricc, and Social Sciences Citation Index". It is available in microform through Univ ersity NMicrofilns, Inc.. 300 North Zeeb Road. Ann Arbor, Nil 48 106, LI.SA. THE WORLD BANK ECONOMIC REVIEW Volume 10 January 1996 Number 1 How Important to India's Poor Is the Sectoral 1 Composition of Economic Growth? Martin Ravallion and Gaurav Datt Is the Debt Crisis History? Recent Private Capital Inflows 27 to Developing Countries Michael Dooley, Eduardo Fernandez-Arias, and Kenneth Kletzer The Surge in Capital Inflows to Developing Countries: 51 An Analytical Overview Eduardo Fernindez-Arias and Peter J. Montiel A SYMPosIuM ON FERTILITY IN SUB-SAHARAN AFRICA Introduction: Fertility in Sub-Saharan Africa 81 Martha Ainsworth The Impact of Women's Schooling on Fertility and Contraceptive 85 Use: A Study of Fourteen Sub-Saharan African Countries Martha Ainsworth, Kathleen Beegle, and Andrew Nyamete Fertility and Child Mortality in C6te d'lvoire and Ghana 123 Kofi Benefo and T. Paul Schultz Contraceptive Use and the Quality, Price, and 159 Availability of Family Planning in Nigeria Bamikale J. Fevisetan anid Martha Ainsworth Fertility, Contraceptive Choice, and Public Policy in Zimbabwe 189 Duncan Thomas and John Maluccio THE WORLD BA NK I,CONNOMI( RFVI-FW, VOL. 1), NO. 1: 1-25 How Important to India's Poor Is the Sectoral Composition of Economic Growth? Martin Ravallion and Gaurav Datt Using a n7ew series of consistenit, consumption-based poverty measures spanning forty years, we assess hou! much India's poor shared in the coun1try's eco7ton7ic groutth, taking into account its urban-rural and output composition. Rural consumption growth reduced poverty in both rural and urban areas. Urblan grouwth brought some beenefits to the urban poor, but had no impact on rural poverty. And rural-to-urban population shifts had no significanit impact on poverty. Decomposing grouwth by output sectors, wve found that output grouth in the primary and tertiary sectors reduced poverty in both urban and rural areas but that secondary sector grouwth did not reduce poverty in either. It is sometimes claimed that the sectoral composition of economic growth is an important determinant of the rate of poverty reduction in developing countries. But testing that claim is difficult. The main evidence cited by those who empha- size the importance of the pattern of growth is a static poverty profile from a single cross-sectional household survey, showing (among other things) where the poor live and the sectors in which they are employed. Poverty profiles for India (and most other developing countries) have indicated higher absolute pov- erty levels in the rural sector. But we cannot automatically assume that rural economic growth is the key to poverty reduction; the rural sector may just not have the potential for high growth. As in most developing countries the trend rate of growth in India has been higher in the modern industrial and service sectors-both of which are mostly urban based-than in the agricultural sector (Chenery and Syrquin 1986). Under certain conditions migration from rural to urban areas may be more important to poverty reduction than rural economic growth (Fields 1980; Anand and Kanbur 1985). The effects of growth in one sector can be crucial to growth in another (Thorbecke and Jung 1994). The fortunes of the poor in each sector are linked-through trade, migration, and transfers-to the living standards of both poor and nonpoor households in other sectors. To avoid small-sample biases in testing the impact of growth on poverty, a reasonable number of time-series observations should be used. But although Martin Ravallion and GaLirav Datt are with the Policy Research Departmeint at the World Bank. The authors would like to thank Lyn Squire, T. N. Srinivasain. Dominique van de Walle, seminar participants at the World Bank, and three anonymous referees for their comments. e 1996 The International Bank for ReconstruCtionI and Development / I [ IF. WORI D) BANK I 2 IHHE WoRI D BANK ECONOMI: REVIEW, V(l.. 1t, NO. I national economic growth can be tracked annually for most countries, the house- hold survey data needed to monitor living standards of the poor are collected much less frequently. Indeed most countries have, at best, a few nationally repre- sentative and comparable surveys spanning a period during which there have been shifts in the sectoral pattern of growth. Among developing countries India has the longest series of national household surveys suitable for tracking living conditions of the poor. At the time of writing distributional data on household consumption in India from thirty-three surveys spanning from 1951 to 1991 could be assembled. The surveys are large enough to be representative at the urban and rural levels as well as nationally, and they are comparable over time because the basic sur- vey method changed very little. There has been much debate about how much India's poor have shared in the country's economic growth. Some critics have argued that the gains in farm output from the green revolution brought little or no gain to the rural poor, while others have pointed to the growth of farm output as the key to reducing rural poverty. (Lipton and Ravallion 1994 review this debate. On the effects of agricultural growth on rural poverty in India see Ahluwalia 1978, 1985; Bell and Rich 1994; Bhattacharya and others 1991; Gaiha 1989; Ravallion and Datt 1994; Saith 1981; and van de Walle 1985.) Views have also differed on how much urban growth has benefited the poor. The optimism of many of India's postindependence planners, who believed that the counltry's (largely urban-based) industrialization would bring lasting, longer-term gains to both the urban and rural poor, has not been shared by many critics (see, for example, Eswaran and Kotwal 1994). And the importance to the poor of the rertiary (mainly services) sector is unclear. Such intellectual debates about growth and the poor lie at the heart of ongoing discussions on development strategy and policy reform in India and elsewhere. (On the role of economic growth on a strategy for poverty reduc- tion see World Bank 1990 and Lipton and Ravallion 1994. On the relevance of these issues to concerns about policy reform and the poor in India see Ravallion and Subbarao 1992.) In this article we report new empirical evidence that sheds light on the effects of the sectoral pattern of economic growth on poverty in India over forty years. We measure the importance to India's poor of initrasectoral growth, rural-to- urban migration, and spillover effects between sectors. We also examine whether these effects differ according to how sensitive the poverty measure is to distribu- tion among the poor. The following section describes how cross-sectoral spillover effects might occur and sets up a framework to test for the effects of sectoral composition and population shifts on poverty during a period of growth. In section 11 we describe our data, including our estimates of a consistent time series of different poverty measures for urban and rural areas of India during 1951-91. Section III then presents our results and discusses their implications. Conclusions are summa- rized in section IV. Ravallion and Datt 3 1. POVERTY AND THE SECTORAL COMPOSITION OF GROWTH Why Would the Sectoral Composition of Growth Matter? Theories of growth and distributional change have emphasized the role played by population shifts from the "traditional" rural sector to the "modern" urban sector. An influential model of this sort was sketched by Kuznets (1955) and later formalized by Robinson (1976), Fields (1980), and Anand and Kanbur (1985), among others. This model attributes growth and distributional shifts to urbanization, assuming that neither mean income nor its distribution changes within each sector. We call this the "Kuznets process" (following Anand and Kanbur). Other strands of the literature have given more attention to intrasectoral changes. Growth of a given sector's output will have a direct effect on incomes of those employed in that sector. In most developing countries the rural sector accounts for a substantially higher share of absolute poverty than the urban sector; a rural resident is also more likely to be poor, by almost anv standard (Lipton and Ravallion 1994 survey the evidence). These stylized facts suggest that the urban-rural composi- tion of economic growth influences poverty reduction. Additionally, there may be indirect cross-sectoral effects arising from the sectoral interdependence of economic activity (Thorbecke and Jung 1994). In principle these may either enhance or retard the direct effect of growth. For the class of additively decomposable poverty measures, national poverty is a population-weighted sum of rural and urban poverty. Thus these measures naturally decompose into population-shift and intrasectoral effects and can illu- minate these issues (Ravallion and Huppi 1991). It is also of interest to look at the relationships among these components. The direct impact of a sector's growth on national poverty is limited by its population share. However, in principle, growth or contraction of one sector can affect other sectors, with potentially wide-ranging implications for poverty reduction. For example, it is often said that an important cause of urban poverty in developing countries is rural pov- erty. According to this view, the vast urban slums of many cities in developing countries are simply the urban analogue of deprivation (often on a larger scale) in the rural hinterland. (For a survey of the literature on poverty in developing countries, including comparisons between urban and rural poverty, see Lipton and Ravallion 1994.) Because of cross-sectoral spillover effects, the significance of the urban-rural composition of growth for poverty extends beyond what is implicit in the sectoral population shares. Spillover effects can occur in a number of ways. Labor mobility between urban and rural sectors can yield an equilibrium relationship between the real wages of similar workers, entailing some degree of horizontal integration in earnings and income distributions; the living standards of similarly endowed people in different sectors are causally related. Even without labor mobility such 4 THE WORLL) BANK ECONOMIC. REVIEW, VOL.. IU, NO. I integration can arise through trade in goods; the living standards of similar house- holds in different sectors will move together to the extent that trade in goods eliminates differences in factor costs at the margin. But even without factor- price equalization, the fact that the rural sector produces food that is consumed in the urban sector means that agricultural growth raises urban welfare by low- ering food prices. Transfers between related households living in different sec- tors can also produce horizontal integration. If the degree of horizontal integration varies with the standard of living, we can also expect growth or contraction in one sector to induce distributional shifts in the other sector. There is no a priori reason to expect the integration to be uniform at all levels. And there is at least one good reason to expect that it will not be: the distributions of living standards in different sectors tend to over- lap imperfectly, that is, they share a positive density over certain (compact) intervals of the range of living standards but not others. For example, the urban sector of a developing country will often include an elite class that simply has no counterpart in the rural sector. This imperfect overlapping can have strong im- plications for how an increase in incomes in one sector will affect both average levels of living and inequalities in other sectors. Testing the Impact of the Sectoral Composition of Growth on Poverty In this section we do not attempt to develop a comprehensive structural model of the potential channels described above. Rather, our aim is to test the impor- tance of the sectoral composition of growth, allowing explicitly for population shifts and cross-sectoral effects. We restrict attention to the broad class of additive poverty measures (Atkinson 1987) and consider two sectors, urban and rural. The average level of poverty is (1) P=n,,P +n,Pr where n1 and P,. are the population shares and povertv measures, respectively, for sectors i = u,r, representing urban and rural areas, respectively. Mean con- sumption can similarly be written (2) = = ps,, + nrpT where p. is the mean for sector i. Let sP= n P /P and s>' = n p/p be the sector shares of total poverty and total consumption income. The growth rate in the poverty measure can be decomposed by taking the total differential of equation 1: (3) dlnP = s,(dlnP,, + ss AInY, + r,'s3AlnY3 + E (12) sFAlnPA = 7r',.sAlnY, +±<2s2ZAInY2 + r3s3A3nYI +Y (13) (sp -sP'nrInu,)Alnnr = it,g,sAlnY, +r,,2s2AInY1 + ',rs3AIn Y3 + E' The breakdown enables us to test for the differential effects of growth in various sectors on urban and rural poverty as well as the effect of rural-to-urban migra- tion. As before, we estimate equations 10-12 and use the condition r*;j = .r - i - 7r* i = 1, 2, 3 to infer the parameters of equation 13. The elasticities of the poverty measures to the sector means can be readily obtained. In the regressions of the national poverty measures (equations 5 and 10) the elasticities are obtained by multiplying the regression coefficients by the relevant consumption or income shares. For the decompositions of the rate of reduction in average poverty (such as equations 7 and 8), the elasticity of pov- erty in sector i (= u, r) to growth in sector j is obtained by multiplying the regres- sion coefficient for i by that sector's consumption or income share relative to i's share of total poverty. II. DATA The Consumption Distributions For this investigation we derived a new and consistent time series of poverty measures for rural and urban India between 1951 and 1991. This time series is based on consumption distributions from thirty-three household surveys con- ducted by the National Sample Survey Organization (NSSO). We use distribu- tions from the third survey round, for August to November 1951, up to the forty-seventh round, for July to December 1991.' This series substantially im- proves upon the most widely used time series on poverty measures in India to 1. The first two rounds of the National Sample Survey (NSS) covered rmral areas only. Ravallion ancI Datt 7 date. Past work has relied on poverty measures presented in Ahluwalia (1978), which gives estimates of the head count index, and Sen's (1976) poverty mea- sure for rural areas, including only twelve rounds, spanning 1956-57 to 1973- 74. One extra round (1977-78) was added in Ahluwalia (1985). Datt (1995) describes in detail how our new series was estimated, so we will be brief here. A set of data discs and a manual are available from the authors. Several points should be made about the consumption distributions. Follow- ing the now well-established practice for India and elsewhere, a household's standard of living is measured by real consumption expenditure per person. The consumption measure is comprehensive, following sound and consistent survey and accounting practices. The underlying NSSO data do not include incomes- although it can be argued that current consumption is a better indicator of living standards than current income.2 Nonetheless, this measure cannot capture vari- ous nonincome dimensions of well-being, and we say nothing here about how responsive these dimensions may be to growth (for further discussion and references see Anand and Ravallion 1993). The average sample size of the thirty-three surveys is 10,988 urban house- holds and 18,691 rural households. But there is considerable variation over time. The urban samples range from 514 to 58,162, while the rural samples range from 1,361 to 99,766. In both cases the smallest samples were in 1953 (al- though in different rounds), and the largest were in 1977-78. From 1955 on- ward all samples exceeded 1,000. We use the urban-rural classification of the NSSO'S tabulations.' Over such a long period some rural areas naturally became urban areas. To the extent that rural (nonfarm) economic growth may foster such reclassifications, it may pro- duce a downward bias in our estimates of the (absolute) elasticities of rural poverty to rural economic growth. The impact on urban elasticities could be positive or negative, depending on the circumstances of the new urban areas relative to the old ones. We have little choice but to use the NSSO'S classification, given that unit record data are unavailable and given that we do not know what the best corrective action would be if we had access to those data. Whenever the dependent and independent variables are estimated from the same survey data, a bias may arise because measurement errors in the survey can be passed on to both variables; if the mean is overestimated, poverty will tend to be underestimated. In all of our regressions we have also tried an instru- 2 Current consumption is a better indicator than currenit incomes particularly in rhis setting. For an overview of supporting arguments see Ravallion ( 1994). Using village panel data from India, Chaudhuiri and Ravallion ( 1994) find that current consumption and income are better indicators of chronic poverty than other measures tested. although the choice between consumption and income is less clear. Even so. it can be argued that current consumption is the better indicator of current level of living. 3. The Nss has followed the census definition of urban areas, which is based on a number of criteria, including a population greater than 5,000, a density niot less than 400 persons per square kilometer, and three-fourths of the male workers engaged inl iioniagriciultuiral pursulIts (Government of India 1992). 8 VHF WORLLD BANK ECONOMIC( REVIEW. VOL. 10, NO. I mental variables estimator, in which the instruments excluded variables derived from the same survey as the dependent variable. The Poverty Line anid Deflators Consistent measurement of absolute poverty requires that the poverty line be the cost of a fixed standard of living over the period of analysis and across sectors (Ravallion 1994). The povertv line we use is the line originally defined by Government of India (1979) and recently endorsed by the Planning Commis- sion (Government of India 1993). This poverty line is based on a nutritional norm of 2,400 calories per person per day in rural areas and 2,100 calories in urban areas. The poverty lines for rural and urban sectors were defined as the level of average per capita total expenditure at which the caloric norms were typically attained in each of the two sectors, following what has been termed the "food energy method" (Ravallion 1994). The rural poverty line was thus set at a per capita monthly expenditure of 49 rupees (rounded to the nearest rupee), and the urban at 57 rupees, measured at 1973-74 prices. The food energy method may not yield consistent poverty lines (representing a uniform threshold in terms of the living stanrdard indicator), especially if the average levels of living vary substantially across sectors (Ravallion 1994). Bet- ter-off regions or sectors will tend to have lower average food shares and hence reach caloric requirements at higher real expenditure levels. This tendency can severelv distort the poverty profile. A case study for Indonesia found that this method produced poverty lines that vary so much in terms of their basic-needs purchasing power that the method produced considerable reranking of regions and sectors (Ravallion and Bidani 1994). However, one can readily test the method for India; independent estimates of the urban-rural cost of living differen- tial can be used in conjunction with the rural poverty line to derive the equiva- lent urban line. For 1973-74 Bhattacharya, Choudhury, and Joshi (1980) esti- mated that the cost of living for the poor was 16 percent higher in urban areas-the same amount (to the nearest integer) implied by the food energy method (although this result may stem from the higher caloric requirement used for rural areas in the Planning Commission's poverty lines).4 It can thus be ar- gued that, for India, the food energy method has not vitiated the urban-rural poverty comparison. After August 1968 the all-India consumer price index for industrial workers (cpilw) is used as the deflator for the urban sector. A detailed discussion of the deflators used for comparisons over time can be found in Datt (1995). We will limit ourselves to only a brief description here. For the earlier period the Labour Bureau's consumer price index for the working class is used, which is an earlier incarnation of the cpinw, albeit with less coverage of urban centers (twenty- seven compared with fifty). The rural cost of living index series was constructed in three parts. For the period since September 1964 the rural cost of living index 4. This is the Fisher index, which gave a differential of 15.9 percent. The Laspeyres index gave 16.5 percenlt, while the Paasche index gave 15.2 percent. Ravallion and Datt 9 is the all-India Consumer Price Index for Agricultural Laborers (CPIAL) pub- lished by the Labour Bureau. For September 1956 to August 1964 (for which an all-India CPIAL does not exist), a monthly series of the all-India CPIAL was con- structed as a weighted average of the state-level CPIALS, using the same state- level weights as those used in the all-India CPIAL published since September 1964. For August 1951 to August 1956 forecasts were obtained from a dynamic model of the CPIAL as a function of the cPiiw and the wholesale price index (for details see Datt 1995). Our CPIAL series also overcame the problem that arose because the Labour Bureau had used the same price for firewood in its published series since 1960- 61. Firewood is typically a common property resource for agricultural laborers, but it is also a market good, and thus the Labour Bureau's practice is question- able. This practice is even more questionable because the NSS values nonpurchased firewood consumption at local market prices (see Minhas and others 1987 for further discussion). Our CPIAL series replaces the firewood subseries in the CPIAL with one based on mean rural firewood prices (only available from 1970) and a series derived by assuming that firewood prices increased at the same rate as all other items in the Fuel and Light category (prior to 1970). The final cpiiw and CPIAL indexes are averages of monthly indexes corre- sponding to the survey period of each of the Nss's rounds. We differ in this respect from Ahluwvalia (1978), who uses averages of the CPIAL over the agricul- tural year (July to June), even for Nss rounds in which the survey period was different. Given the seasonality of prices, matching the survey period is arguably a better procedure. The National Accounts and Population Data Our data on sectoral incomes are taken from various issues of the National Accounts Statistics (NAS) published by the Central Statistical Organization (cso). In particular, we draw upon the NAS to create an annual series of the net domes- tic product (NDP) at factor cost at constant 1980-81 prices, and its sectoral components, that is, NDP in the primary, secondary, and tertiary sectors. The constant price conversions implicit in these series are based on the national ac- counts deflators. The primary sector includes agriculture, forestry, fishing, min- ing, and quarrying; the secondary sector includes manufacturing, construction, and electricity, gas, and water supply; the tertiary sector includes trade, hotels, restaurants, transport, storage, communication, finance, insurance, real estate, business services, and community, social, and personal services. We also draw upon the NAS to construct a series on private final consumption expenditure at constanit prices as an alternative to the Nss-based series. The NAS reports these series annually for the financial year April to March. To mesh these data with the poverty data from the NSSO, we linearly interpo- lated the annual national accounts data to the midpoint of the survey period for different rounds. But the first ten NSs rounds covered periods shorter than one year (from four to nine months), and thus the mapping into annual national I 0 THE WORLD BANE ECONOMIC RiVIEW, VOl. 1(, NO. I accounts data was far more problematic. We thus deleted the first ten surveys in any regressions using national accounts data. The population estimates are based on the census population totals and as- sume a constant growth rate between censuses. They are also centered at the midpoints of NSS survey periods. Poverty Measures We use three poverty measures: * The headcount index, given by the percentage of the population that lives in households with a per capita consumption below the poverty line. * The poverty gap index, defined by the mean distance below the poverty line, expressed as a proportion of that line. The mean is formed over the entire population, counting the nonpoor as having a zero poverty gap. * The squared poverty gap index introduced by Foster, Greer, and Thorbecke (1984), defined as the mean of the squared proportionate poverty gaps. Unlike the poverty gap index this measure is sensitive to distribution among the poor. A transfer of income from a poor person to a poorer person, for example, will not alter either the headcount index or the poverty gap index, but it will decrease the squared poverty gap index. Furthermore-and unlike the Sen (1976) distribution sensitive measure of poverty-the squared poverty gap index satisfies the "subgroup consistency" property; that is, if poverty increases in any subgroup (say the urban sector), and it does not decrease elsewhere, then aggregate poverty must also increase (Foster and Shorrocks 199 1). All three measures are members of the Foster-Greer-Thorbecke (FGT) class, for which the individual poverty measure is (14) = (I - x, /z)( if xi < z = 0 if x,> z where x. is consumption expenditure of the ith person in a population of size n, z is the poverty line, and a is a nonnegative parameter. Average poverty is simply (15) Py = pa ,/71. The headcount index is obtained when a= 0, the poverty gap index when a= 1, and the squared poverty gap index when a = 2. The poverty measures are estimated from the grouped data on consumption distributions using paramet- erized Lorenz curves (see Datt and Ravallion 1992). III. RESULTS The urban sector's share of consumption has risen steadily since about 1960 (figure 1). Both the secondary and tertiary sectors' shares of national income Ravallion and Datt 11 Figure 1. Sectorail Composition of Economic Actifuity in India, 1951-91 Percent 40 Tertiary sector 35 Urban areas 30 - 2 _ ~ ~ ~ ~ ~ ~ ~ ~ ~~ S~~~econdary sector 259 20 -, 15- 10 - I I I I 1990 1955 1960 1965 197( 1975 1980 1985 1990 A'otc The urban share is of total national constimption as estimated from National Sample Surveys. The secondary and tertiarv shares are of net domestic product as estimated from the National Accounts Statistics. Souirce Authors calculations from National Sample Surveys and National Accounts Statistics. have been on a trend increase over the whole period (the balance is the primary sector). There have been sizable fluctuations in the NSSO-based means of consump- tion, although some patterns are evident (figure 2). There was a contraction in the early 1950s, followed by a long period of stagnation, and a reasonably sus- tained period of growth since the mid- 1970s. Throughout the entire period there is strong comovement between the urban and rural means (the simple correla- tion coefficient is 0.84; the correlation coefficient of the first differences be- tween survey rounds is 0.49). Thus the historical gap in average living standards between the sectors was maintained: there is no significant time trend in the ratio of the rural to the urban mean.' The consumption mean derived from the national accounts shows a reasonably strong trend increase over the whole period, and has been higher than even the urban Nsso-based mean since the mid-1960s. The discrepancy between the national accounts consumption numbers and those from the NSSO has been noted before, and we will not discuss the issue here. For 5. If the log of the ratio of the rireans is regressed on time and one corrects for the serial correlation in the errors, the implied rate of growth in the ratio of the urban mean to the riral mean is -1.4 percent per year, but the t-ratio is only 1.2. 1 2 THE WORI I) BANK CONOMI(. RlVIEW. VOI. 11, No. I Figure 2. Average Conszmption in India, 1951-91 Mean consumption (rrupees per month per person; 1973-74 prices) 110 100 90- NAS 80- 70 - NSS-urban 60 - S0-\ I % /e %, ,0 NSS-rural 50 - ~ ~ ~ ' 40- l l l l l l 1950 1959 1960 1969 1970 1975 1980 1985 1990 Source.- ALithors c1lculations from National Sample Surveys (NSS) and National Accouints Statistics (NAS). further discussioni see Vaidyanathan (1986), Suryanarayana and Iyengar (1986), Minhas (1988), and Bhattacharya and others (1991). The bulk of the consumption growth since about 1970 is attributable to growth within sectors; the Kuznets process of rural to urban migration, at given sector means, accounts for very little. Averaging the three survey rounds in 1969-71 and the three rounds in 1989-91, we find that only 6.4 percent of the increase in log consumption from 1970 to 1990 is attributable to population shifts (the third term in equation 4), while 20.0 and 73.5 percent are attributable to growth within the urban and rural sectors, respectively (given the initial urban population share). Neither the headcount index nor the squared poverty gap index for either urban or rural sectors shows a trend increase or decrease until about 1975, when a trend decrease emerged (figure 3). The pattern of change over time is very similar for the poverty gap index (see Datt 1995 for details). This pattern also holds for urban poverty although the fluctuations seem less pronounced. Comovement is strong: the simple correlation coefficient between the contem- poraneous sector values of the log headcount index is 0.92 (0.68 between the first differences). There are also signs of convergence between urban and rural areas by the end of the period, with the urban squared poverty gap overtaking the rural index. However, the rural sector still accounts for 74 percent of the Ravallion and Datt 13 Figure 3. PovertyMeasuresforIndia, 1951-91 Headcount index (percent) Squared poverty gap (x1OO) 70 - 25.0 A - 22.5 60 - I Rural headcouL1nt -20.0 40 - I/ _ _ > - 17.5 19.0 40-IN Uxrban headCOLunt - 12.5 30' 10.0 5.0 10 - Urban squarecl povertv gap -2.5 0)- I I I IIIII -0 1950 19 1960 1965 1970 1975 198( 1985 1990 Source: AuLIorsB calculations. total number of poor at the end of the period, 70 percent of the average poverty gap index, and 68 percent of the average squared poverty gap index. As with growth in mean consumption, the bLIlk of the poverty reduction after 1970 is attributable to gains within sectors rather than the population shift ef- fect. For example, the impact of population shifts accounts for only 3.2 percent of the difference in the log headcount index between 1970 and 1990, while the urban and rural sectors account for 12.3 percent and 84.5 percent, respectively (the results are similar for the squared poverty gap). The Growth Elasticities of National Poverty Measuires The elasticities of all three poverty measures with respect to the three mea- sures of economic growth-the mean consumption per person as estimated by the NSSO, the mean consumption per person as estimated by the national ac- counts and population census, and the mean NDP ("income" for short) per per- son also taken from the national accounts and census-are estimated by regress- ing the first difference of the log poverty measure against the first difference of log mean consumption or income (table l). We also give an "adjusted" estimate in which another variable was added, namelv the first difference of the log of the ratio of the consumer price index for agricultural laborers to the national in- Table 1. Elasticities of National Poverty Measures to Economi2ic Growth in India Elasticity tith respect to Mean consumption Mean private consumption from fronm national national accounts Mealan net domestic product Poverty measure sample surveys Unadjusted Adjustedt Unadjuisted AdjustedJ Headcount index (a = 0) -1.33 -1.21 -0.90 -0.99 -0.75 (15.19) (4.04) (4.23) (3.38) (3.68) Poverty gap index (a = 1) 1.88 -1.79 -1.36 -1.49 -1.15 (12.83) (4.02) (3.98) (3.44) (3.59) Squared poverty gap index (a = 2) -2.26 -2.18 -1.67 -I.XS -1.45 (10.22) (3.73) (3.45) (3.32) (3.2 ! Note: Absolute t-ratios in parenthieses. Based on regressions of first differences of rhe log povei ry measuires against fi rst differences of the log consumption or net product per person, u sing thirty-three surveys spanninig 1951-91 for estimating the elasticity with respect to the surveys-based mean consumption and rwentv-three surveys Spallnlinlg 1958-91 for estimating elasticities to consumption or incomne from the national accounits. All regressions comfortably passed residual diagnostic tests for serial correlation, functional form, normality, and heteroskedasticity (see appendix for details). a. The adjusted estimates include an additional regressor, that is the difference in the rates of inflation implied by the consumer price index and the national income deflator. Souirce: Authors' calculationis. Ravallion and Datt 15 come deflator (that is, the difference in the rate of inflation implied by the two deflators). This variable was included to allow for possible bias in estimating the growth elasticity, which may arise because of the difference in the deflator used for the national accounts data and that used for the poverty lines. The national poverty measures responded to all three measures of economic growth. The elasticities are higher if the NSSO estimate of mean consumption is used rather than the national accounts estimate, although the difference is not large for a given value of a. The elasticities are lowest for per capita NDP. This result may be due to intertemporal consumption smoothing, which may make poverty (in terms of consumption) less responsive to income growth than to consumption growth in the short term. The Impact of the Urban-Rural Comzposition of Growth on Poverty All of the regressions of equations 5, 7, and 8 fitted well and passed almost all standard tests on the residuals (table 2). However, a correction for serial corre- lation in the residuals was needed in some of the urban poverty equations. The appendix tables give results for the complete regressions and various statistical tests. Table 2. Impacts of the Llrban-RRural Composition of Grouwth on Poverty in India Poverty measure National pouertv I irban potvertv Rural poverty Urban growth Headcount -0.549 -0.560 -0.169 (1.367) (5.687) (0. 542) 1-0.1421 1-0.8241 1--(O.3I Poverry gap -0.288 -0.(.23 0.278 (0.449) (4.821) (0.436) 1--0.0751 1-0.9151 10.0871 Squared poverty gap 0.234 -0.5l59 0.77 (0.244) (3.302) (0.805) (0.0611 1-(.8291 1((.243] Rural growth Headcount -1.461 -0.076 -1.141 (12.636) (2.951) (I2.69 1 1-1.0831 [-0.320)1 1-1.0271 Poverty gap -2.123 -0.129 -1.979 (11.502) (3.9477) (10.800) 1-1.574] 1-0.5431 1-1.7X1 Squared poverty gap -2.651 -0.174 -2.446 (9.585) (4.11 3) (8.8 11) 1-1.9651 [-0(7391 1-2,1971 Note: Absolute t-ratios in parentheses; elasticities at mean points in brackets (see appendix tables for detailed results). At the mean points, the urban share of national cosumiption was 0.259, and the priliary- secondary-tertiary brealkdowin was 0.487, 0.202, 0. 312. he urhail share of total posertv was 0. 176 for both the headcount index and the poverty gap index and 0.175 for the squared poverty gap. Source: Authors' calculations. 16 THE WORIL) BANK FCONOMN11R REVIEW. VOL 10, NO. I There is strong evidence that the urban-rural composition of growth matters to India's rate of progress in reducing all three national poverty measures. The urban growth effect is not significantly different from zero in explaining the rate of poverty reduction nationally. But the rural growth term is highly significant. A Wald test of the null hypothesis that urban consumption growth has the same effect on national poverty as rural growth can be rejected in all cases (appendix tables A-t, A-2, and A-3). A stronger version of the test requiring uniform ef- fects of urban and rural growth as well as sectoral population shifts was also rejected for the poverty gap and the squared poverty gap indexes, though we were unable to reject it for the headcount index (see the appendix tables for details). We foulld that the urban-rural population shift had no significant ef- fects on poverty. Thus it appears that the strong growth effects evident in table 1 are largely attributable to rural consumption growth, with very little contribu- tion from either urban growth or the Kuznets process. Turning to the urban-rural decomposition of poverty reduction, we find that urban growth reduced urban poverty (table 2), but so did rural growth, which had a significant impact on poverty in both sectors for all three poverty mea- sures. Indeed for the squared poverty gap the elasticity of urban poverty to rural growth is almost as high as it is to urban growth. The effect of urban growth is too small to be detected in the national average poverty measures. The relatively low impact of urban growth on urban poverty and the propoor spillover effect of rural growth suggest significant distributional effects within urban areas. This proposition is confirmed by regressing the change between surveys in the (log) Gini index for urban areas on the growth rates in both urban and rural means:6 (16) A In Ginii(urbani) = 0.73A In mnean(urban) - 0.41A In ntean(rural). (5.85) (4.15) Urban consumption growth has been increasing inequality in urban areas, while rural growth has improved urban distribution. By contrast, performing the same regression for the rural Gini, the changes in either mean are not significant, either individually or jointly. The Impact of the Output Composition of Grou.!th on Poverty In explaining the rate of progress in reducing poverty nationally, we find a marked difference between the primary and tertiary sectors on the one hand and the secondary sector on the other (table 3). Growth in both the primary and tertiary sectors was poverty reducing, the tertiary sector generating a larger impact, though the difference between these two sectors is not significant. By contrast, growth in the secondary sector had an adverse impact, though not significantly different from zero at the 5 percent level for any poverty measure. 6. Absolute t-ratiosaregivxenin parentheses.TheR-is 0.65 .Acorrecrion wasimai(leforserial correlation in the error>. Ravallion and Datt 17 Table 3. Impacts of the Output Composition of Growth on Poverty in India Sector and poverty measure National poverty Urban poverty Rural poverty Primary sector growth Headcount -1.157 -0.316 -0.858 (2.964) (2.755) (2.625) [-0.563] [-0.8721 [-0.507] Poverty gap -1.586 -0.432 -1.313 (2.615) (2.983) (2.341) 1-0.772] l-1.1941 [-0.776] Squared poverty gap -1.905 -0.471 -1.660 (2.192) (2.719) (2.040) 1-0.927] [-1.3131 [-0.9791 Secontiary sector growth Headcount 3.409 0.609 2.531 (1.837) (1.176) (1.629) [0.688] [0.697] [0.620] Poverty gap 5.816 1.254 5.162 (2.016) (1.917) (1.936) 11.174] 11.437] 11.265] Squared poverty gap 7.026 1.532 6.338 (1.700) (1.960) (1.637) [1.4181 [1.771] [1.550] Tertiary sector growth Headcount -3.418 -0.702 -2.373 (2.737) (2.009) (2.270) [-1.065] [-1.240] [-0.898] Poverty gap -5.869 -1.216 -5.124 (3.024) (2.755) (2.856) [-1.1851 [-2.151] [-1.938] Squared poverty gap -7.274 -1.458 -6.449 (2.616) (2.763) (2.476) [-1.468] 1-2.601] [-2.434] Note: Absolute t-ratios are in parentheses; elasticities at mean points in brackets. These regressions were augmented for differences in the rate of inflation implied by the consumer price indexes (CPis) and national income deflators (see appendix tables for details). Source: Authors' calculations. When we turn to the tests for output compositional effects on the rates of poverty reduction in the urban and rural sectors we find that primary and ter- tiary sector growth was poverty reducing in both urban and rural sectors (table 3). The highest elasticities were for tertiary growth, although tertiary growth started from a smaller base. By contrast, secondary sector growth had no signifi- cant impact on the rate of poverty reduction in either urban or rural areas. One striking feature of the results in table 3 is that, for a given poverty mea- sure, the coefficients on secondary and tertiary growth components are nearly equal in absolute value, although they have opposite signs. In fact, the null r3 = -r, cannot be rejected at the 1 percent level for any poverty measure, and this result also holds for the urban-rural components of the change in poverty (ap- 18 IHE WORLD BANK E C)N0M[( Rl:VII'W, VoL. 1(0, NO. I pendix table A-4). In terms of equation 10, this result is telling us that the rela- tionship can be simplified to (17) AInP=,r1s1AlnY1 + r3A(Yi -Y2) +£ and our regressions indicate that r, and 7r, are negative. Thus it is not tertiary sector growth per se that is reducing poverty, but increases in the difference between tertiary sector output and the output of the (smaller) secondary sector. At first sight this result seems odd, but it has a plausible interpretation. In India (as in other developing countries) the tertiary sector includes a hybrid of activities that are of varying importance to the poor. It combines, for example, formal sector finance and insurance firms with informal trade and transport ac- tivities. Let Y3 - Y3 + Y3V, where the subscripts 3/and 3i refer to the formal and informal tertiary sectors, respectively. Suppose that the true relationship between the rate of progress in reducing poverty and the composition of growth is (18) AInP = rTlsAlnY, + 7r,l, n ,AY3 + gY. This equation assumes that (in addition to the primary sector) it is the informal tertiary sector that matters to the poor, and not the secondary sector or the formal tertiary sector. However, the formal tertiary sector is likely dependent on the secondary sector: secondary sector growth generates demand for out- puts from the formal tertiary sector. Suppose, in particular, that average in- comes in the secondary sector move so closely to those in the formal tertiary sector that the unobserved variable AY3f is approximated well by AY,. Under these assumptions, equation 1 8 implies that equation 17 is an estimable model. According to this interpretation, the secondary sector is acting as a proxy for the formal tertiary sector, when in fact it is growthi in the informal tertiary sector that matters to the poor. Does the Poverty Measure Matter? Tables 1-3 show that our qualitative results are robust to the choice of pov- ertv measure. However, the growth elasticities tend to be higher (in absolute value) for higher values of a. To help interpret this resulr, note that the poverty gap indexes can be written in a nested form: Pi = P(I-p(Iz) ( 19) PI = Pi[1 + '/z+ +(aPzj1 where puP and uP are the mean and standard deviation of consumption by the poor, respectively. As can be seen from equation 19, the higher growth elasticity of P1 compared with P( implies that the depth of poverty (as measured by 1 - ,uPlz) is also reduced by growth. Similarly, the higher elasticity of P, relative to Pi implies that inequality among the poor-as measured by the coefficient of Ravallion and Datt 19 variation-is reduced by growth. Thus the effects of growth within and between sectors are not confined to households within a neighborhood of the poverty line. IV. CONCIUSIONS Despite the substantial sectoral shifts that have occurred over the last forty years, poverty in India is still overwhelmingly rural. At the beginning of the 1990s, 74 percent of the country's poor lived in rural areas. That fact alone does not imply that urban economic growth is unimportant. The nature of intrasectoral and intersectoral effects of growth and of rural-to-urban migration on poverty may mean that rural economic growth is far less important than the sheer size of the rural sector would suggest. In fact, the main conclusion of this article holds that, if anything, the opposite is true: the relative effects of growth withiln and between eachl sector reinforced the importance of rural economic growth to national poverty reduction in India. Both the urban and rural poor gained from rural sector growth. By contrast, urban growth had adverse distributional effects within urbani areas, which mili- tated against the gains to the urban poor. And urban growth had no discernible impact on rural poverty. Nor did the (much researched) Kuznets process of growth through rural-to-urban migration significantly reduce poverty in India. When we decompose growth in national income by sectors defined by out- put, we again find marked differences in the impact on poverty. Both primary and tertiary sector growth reduced poverty nationally and within urban and rural areas. By contrast, secondary sector growth had no discernible positive effect on the poor in either urban or rural areas. In the historical shift from the primary sector to the secondary and tertiary sectors it seems that it was the tertiary sector that delivered significant gainis to India's poor. Our investigation points clearly to the quantitative importance of the sectoral composition of economic growth to poverty reduction in India. Despite the rising urbanization of Indian poverty, it is likely to remain true for many years to come that-from the point of view of India's poor-it is the dog (the rural economy) that wags the tail (the urban sector), not the other way around. Fostering the conditions for growth in the rural economy-in both primary and tertiary sectors-must thus be considered cenitral to an effective strategy for poverty reduction in India. But there is another more subtle implication for the future. We have studied the histori- cal experience in a period in which India's development strategy (starting from the Second Plan in the 1950s) emphasized capital-intensive industrialization concen- trated in the urban areas of a largely closed econony. It ma' not be surprising that urban economic growth fueled by such industrializationi brought negligible gains to the poor. This result underlines the importance of making a successful transition to an alternative industrialization process; even then (we suspect), the tail will not wag the dog. But it could surely do a lot more to help it move. 7. Note that a higher growth elasticiry for P, compared with P, implies that *U"L must he increasing in p. and thus a higher elasticity for P, relative to PI muiu imply that &' is decreasing in p. 20 THE WORLD RANh ECONOM1IC REVIEW, VOL. I)0 NO. I Table A-1. Change inz Headcount Index as a Function of Urban and Rural Consumption Growth in India Cbange in Componients of change in national poverty national povlerty Urban Rural Variable or statistic- 01.S Iv AR(]) I v OLS IV Urbao growth (irn) --0.549 -0.445 -0.560 -0.489 -0.169 -0.185 (1.37) (0.937) (5.69) (4.29) (0.54) (0.50) Rural growth (7rm, -1.461 -1.498 -0.076 -0.087 -1.141 -1.154 (12.64) (11.00) (2.95) (2.67) (12.69) (10.91) Population shift (rj) -4.458 -4.718 -0.775 -0.908 -1.624 -1.534 (1.31) (1.35) (1.25) (1.08) (0.61) (0.56) R' 0.895 0.894 0.761 0.732 0.886 0.886 Standard error of estimare 0.0295 0.0295 0.0068 (.0071 0.0229 0.0229 Autocorrelation (1) 2.555 2.494 n.a. 3.161 1.812 2.026 Functional form (I) 0.006 0.005 n.a. 0.191 0.071 0.062 Normalitv (7) 0.059 0.056 n.a. 0.613 0.552 0.412 Heteroskedasticity (1) 0.132 0.157 n.a. 0.224 0.634 0.544 Sargan's IV test (11) n.a. 4.765 n.a. 4.903 n.a. 5.926 Wald test (2): r=lt,=7t1 4.10 3.881 nIa. n.a. n.a. n.a. Wald test (1): ,r1=yr, 3.879 3.656 n.a. n.a. n.a. n.a. n.a. Nor applicable. Note: The table shows ordinary least squares (ot s) anld instrumental variables (IV) parameter estimates for equations 5, 7, and 8. Absolute t-ratios are given in parentheses. The data are from thirty-three houselhold surveys spanninig from 1951 to 1991 (see text for details). The equations for urban poverty correct for first-order autocorrelation, AR( 1 ). The following set of instruments were used in the IV estimation: date (midpoint) of the survey; time interval betweeni the surveys; lagged rural and urban log real mean consumption; currenit rural and urban price indices (in logs) and their lagged values; change in log real per capita outpILt from the primary, secondary, and tertiary sectors, and log real per capita consumption from the national accounts and its lagged valoe. The bottom part of the table reports a number of diagnostic rests; the test statistics are distributed asX' with the degrees of freedom as noted in parentheses. The last two rows report Wald tests on the null of n[o comiipositional effects of growth on poverty. The stronger version tests the restrictioll that the effects on poverty of urban growth, rural growth, and sectoral population shift are the same; the weaker version tests for uniform effects of urban and rural growtth onl). Souirce: Autrhirs' Calculations. Ravallion anrd Datt 21 Table A-2. Change in the Poverty Gap Index as a Function of Urban and Rural Consumption Growth in India Change in Components of change in n.ational pov'crty nation,al poverty Urbanr Rural Variable or statistic OLS IV AR(l) IV/AR(1) OLS IV Urban growth (ir,) -0.288 -0.116 -0.623 -0-.534 0.278 0.399 (0.45) (0.1 53 (4. 8) ().9() (0.44) (0.53) Rural growth (it,) -2.123 -2.157 -0.129 -0.152 -1.979 -2.029 (11.50) (9.92) (3.98) (3.52) (10.80) (9.40) Population shift (n,) -9.284 -9.555 -1.327 - 1.463 -7.59S -7.524 (1.71) (1.71) (1.82) (1.69) (1.41) (1.35) R2 0.868 0.868 0.777 0.769 0.841 (0.841 Standard error of estimate 0.0471 0.0471 0.0092 (.0095 0.0467 0.0468 Autocorrelation (1) 1.314 1.194 na. na. 1.300 1.350 Functional form (1) 0.648 0.635 n.a. n.a. 1.708 1.730 Normality (2) 0.088 0.105 nO. n.j. 0.463 0.528 Heteroskedasticity (1) 0.324 0.391 n.j . na. 0.006 0.020 Sargan's IV test (11) n.a. 9.968 n.a.. 5.551 n.a. 11.398 Wald test (2): ir1=,t2=ir, 6.813 6.021 na.. n a. n .a. n.a. Wald test (1): 7rl=,r, 6.165 5.399' n.a. n .. n.a. n.a. n.a. Not applicable. Note: See note to table A-i. Source: Authors' calculations. 22 THF WORI D BANK ECONOMI(C RFVIFW. VOL. II). NO. I Table A-3. Chanige in the Squared Poverty Gap Index as a Function of Urban and Rural Consumption Growth in India Change in Components of change in national poverty national poverty Urban Rural Variable or statistic OLS IV AR(l) IV/AR(I) OLS IV Urban growth (ri) 0.234 0.258 -0.559 -0.558 0.777 0.786 (0.24) (0.23) (3.30) (2.29) (0.44) (0.69) Rural growth (i1,) -2.651 -2.649 -0.174 -0.185 -2.446 -2.476 (9.59) (8.14) (4.11) (3.2.3) (8.81) (7.58) Population shift (ir, -13.578 -13.301 -1.924 -1.783 -11.733 -11.044 (1.67) (1.59) (2.06) (1.56) (1.44) (1.31) R2 0.811 0.811 0.748 0.738 0.771 0.771 Standard error of estimate 0.0705 0.0705 0.0122 0.(0126 0.0708 0.0708 Autocorrelation (1) 0.681 0.622 n.a. n.a. 0.561 0.585 Functional form (1 0.399 0.431 n.a. n.a. 1.265 1.366 Normality (2) 0.250 0.256 n.a. n.a. 0.275 0.292 Heteroskedasticity (I) 1.395 1.447 n.a. n.a. 0.182 0.121 Sargan's IV test (1) n.a. 10.105 n.a. 5.928 n.a. 11.024 Wald test (2): zra=r1=,r1 7.434 5.474 n.a. n.a. n.a. n.a. Wald test (1): ir,=r. 6.790 4.889 n.a. n.a. n.a. n.a. n.a. Not applicable. Note: See note to tahle A- I. Source : Authors' calculations. Table A-4. Change in Poverty as a Function of the Primary-Secondary-Tertiary Composition of Growth in India Headcount index Poverty gap index Squared poverty gap index Components of Compottents of Components of Change in change in national Change in change in national Change in change in national national poverty national poverty ttational poverty Variable or statistic povertv UJrban Rural poverty Urban Rural poverty Urban Rural Primary sector growth (itr) -1.158 -0.316 -0.858 -1.586 -0.432 -1.313 -1.905 --0.471 -1.660 (2.96) (2.76) (2.62) (2.62) (2.98) (2.34) (2.19) (2.72) (2.04) Secondary sector growth (it,) 3.409 0.609 2.531 5.816 1.254 5.162 7.026 1.532 6.338 (1.84) (1.18) (1.63) (2.02) (1.92) (1.94) (1.70) (1.96) (1.64) Tertiary sector growth (t3) -3.418 -0.702 -2.373 -5.869 -1.216 -5.124 -7.274 -1.458 -6.449 (2.74) (2.01) (2.27) (3.02) (2.75) (2.86) (2.62) (2.76) (2.48) Change in rural price index 0.939 0.726 1.284 1.173 1.512 1.409 relative to NDP deflator (5.63) (5.20) (4.96) (4.90) (4.07) (4.05) Change in urban price index 0.254 0.350 0.467 relative to NDP deflator (2.71) (2.95) (3.30) R2 0.752 0.491 0.701 0.714 0.490 0.699 0.631 0.469 0.618 Standard error of estimate 0.0378 0.0095 0.0316 0.0587 0.0120 0.0543 0.0841 0.0144 0.0788 Autocorrelation (1) 0.143 0.002 0.158 0.147 1.024 0.001 0.273 4.084 0.101 Functional form (1) 2.813 0.191 2.762 5.036 0.086 3.513 5.953 0.109 3.975 Normality (2) 1.300 0.940 0.527 1.677 0.123 0.924 1.547 0.516 0.059 Heteroskedasticity (1) 0.060 1.693 0.145 0.226 0.075 0.820 0.297 0.018 0.669 Wald test (2): ir= = =2 Kt3 5.109 n.a. n.a. 6.065 n.a. n.a. 4.468 n.a. n.a. Wald test (2): i1rl = it3, ir2 = 0 3.731 n.a. n.a. 5.925 n.a. n.a. 4.057 n.a. n.a. Wald test (1): it2 + it3 = 0 0.000 0.199 0.054 0.002 0.020 0.001 0.019 0.055 0.004 n .a. Not applicable. Note: The table gives least squares parameter estimates for equations 10- 12, n 23. The associated absolute t-ratios are given in parentheses. The bottom part of the table reports a number of diagnostic tests. The test statistics are distributed as x2 with the degrees of freedom as noted in parentheses. The last two rows report Wald tests on the nulls of (a) no sectoral composition effects of growth on poverty (testing for uniform effects of primary, secondary, and tertiary sector growth); (h) the effects of primary and tertiary sector growth are the same, but secondary sector growth has a zero effect; and (c) the effects of secondary and tertiary sector growth are equal but of opposite signs. Source: Authors' calculations. 24 rHl W\C)RLD BANK EC:ONO(MI REVIEW, VOLI. 1. No. I REFERENCES The word "processed" describes informally reproduced works that may not be com- monly available through library systems. Ahluwalia, Montek S. 1978. "Rural Poverty and Agricultural Performance in India." Journal of Development Studies 14:298-323. 1985. "Rural Poverty, Agricultural Production, and Prices: A Reexamination." In J. W. Mellor and G. M. Desai, eds., Agricu/ltural Change and Rural Poverty. Baltimore: Johns Hopkins University Press. Anand, Sudhir, and Ravi Kanbur. 1985. "Poverty under the Kuznets Process." The Economic Journal 95(supplement):42-50. Anand, Sudhir, and Martin Ravallion. 1993. "Humann Development in Poor Countries: The Role of Private Incomes and Public Services." journal of Economic Perspectives 7:133-50. Atkinson, Anthony B. 1987. "On the Measurement of Poverty." Ecoizometrica 55:749-64. Bell, Clive, and R. Rich. 1994. "'Rural Poverty and Agricultural Performance in Post- Independence India." Oxford Bulletin of Economics and Statistics 56(2): 111-33. Bhattacharya, N., D. Coondoo, P. Maiti, and R. Mukherjee. 1991. Poverty, Inequality and Prices in Rural India. New Delhi: Sage Publicationis. Bhattacharya, S. S., A. B. Roy Choudlhury, and P. D. Joshi. 1980. "Regional Consumer Price Indices Based on NSS Household Expenditure Data." Sarvekshana 3:107-21. Chaudhuri, Shubham, and Martin Ravallion. 1994. "How Well do Static Indicators Identify the Chronically Poor?" journal of Public Economics 53:367-94. Chenery, H. B., and Moshe Syrquin. 1986. "Typical Patterns of Transformation." In H. B. Chenery, S. Robinson, and Moshe Syrquin, eds., Industrialization and Grotvth. New York: Oxford University Press. Datt, Gaurav. 1995. "Poverty in ildia, 1951-1992 Trenids and Decompositions." World Bank, Policy Research Department, Washingtoin, D.C. Processed. Datt, Gaurav, and Martin Ravallion. 1992. "Growth and Redistribution Components of Changes in Poverty Measures: A Decomposition with Applications to Brazil and India in the 1980s." /ournal of Development Economics 38:275-95. Eswaran, Mukesh, and Ashok Kotwal. 1994. Why Poverty Persists in India? New Delhi: Oxford University Press. Fields, Gary. 1980. Poverty, Inequality an(d Developm11en?t. Cambridge, Mass.: Cam- bridge University Press. Foster, James, and Tony Shorrocks. 1991. "Subgroup Consistent Poverty Indices." Econometrica 59:687-709. Foster, James, J. Greer, and Erik Thorbecke. 1984. "A Class of Decomposable Poverty Measures." Economietrica 52:761-65. Gaiha, Raghav. 1989. "Poverty, Agricultural Production and Price Fluctuations in Ru- ral India: A Reformulation." Cambridge Joturnal of Economics 13(2):333-52. Government of India. 1979. Report of the Task Force on Projections of Minimumul Needs and Effective Consumnption. New Delhi: Planning Commission. 1990. "Country Paper on the Indian National Sample Survey." National Sample Survey Organizationi. Ministry of Planning, Departmenit of Statistics, New Delhi. Processed. Ravallion and Datt 2i . 1992. Final Population Totals: Brief Anialysis of Primary Census Abstract. Paper 2. New Delhi: Registrar General and Census Commissioner. . 1993. Report of the Expert Group on Estimation of Proportion and Nuimber of Poor. New Delhi: Planning Commission. Kuznets, Simon. 1955. "Economic Growth and Income Inequality." Americani Eco- nomfic Reviewu 45(1): 1-28. Lipton, Michael, and Martin Ravallion. 1994. "Poverty and Policy." In Jere Behrman and T. N. Srinivasani, eds., Handbook of Development Economics. Vol.3. Amsterdam: North-Holland. Minhas, B. S. 1988. "Validation of Large-Scale Sample Survey Data: Case of NSS Esti- mates of Household Consumption Expenditure." 'Sankbya 50(Series B, Pt. 3, Supple- ment): 1-63. Minhas, B. S., L. R. Jain, S. M. Kansal, and M. R. Saluja. 1987. "On the Choice of Appropriate Consumer Price Indices and Data Sets for Estimating the Incidence of Poverty in India." Indian Economnic Review 22((1): 19-49. Ravallion, Martiii. 1994. Poverty Comparisons. C.hur, Switzerland: Harwood Academic Press. Ravallion, Martin, and Benu Bidani. 1994. "How Robust Is a Poverty Profile?" The World Bank Economic Review! 8(l):75-102. Ravallion, Martin, and Gaurav Datt. 1995. "Growth and Poverty in Rural India." wirs 1405. World Bank, Policy Research Department, Washington, D.C. Processed. Ravallion, Martin, and Monika Huppi. 1991. "Measuring Changes in Poverty: A Meth- odological Case Study of Indonesia During anl Adjustment Period." The World Bank Economic Review 5(1 ):57-84. Ravallion, Martin, and Kalanidhi Subbarao. 1992. "Adjustment and Human Devel- opment in India." Journal ot Indian School of Political Economny 4(1):55-79. Robinson, Sherman. 1976. "A Note on the ti-Hypothesis Relating Income Inequality and Economic Developnmenit." American Economic Review 66:437-40. Saith, Ashwani. 1981. "Production, Prices and Povterty in Rural India." journal of De- velopment Studies 19:196-214. Sen, Amartya. 1976. "Poverty: An Ordinal Approach to Measurenmenit." Econometrica 46:437-46. Suryanarayana, M. H., and N. S. lyengar. 1986. "On the Reliability of NSS D)ata." Economic and Political Weekly 21:261-64. Thorbecke, Erik, and Hong-Sang Jung. 1994. "A Multiplier Decomposition Method to Analyze Poverty Alleviation." Cornell University, Department of Economics, Ithaca, N.Y. Processed. Vaidyanathan, A. 1986. "On the Validity of NSS Consumption Data." Economic and Political Weekly 21:129-37. van de Walle, Dominique. 1985. "Population Growth and Poverty: Another Look at the Indian Time Series Data." journal of Development Studies 21:429-39. World Bank. 1990. World Development Report 1990: Poverty. New York: Oxford University Press. Is the Debt Crisis History? Recent Private Capital Inflows to Developing Countries Michael Dooley, Eduardo Fernrndez-Arias, and Kenneth Kletzer The ouitlook for economic development for an important group of middle-income countries has once again been buoyed by substantial private capital inflows in the 1990s. As in the 1970s, this development has been met with cauitious optimism. This empirical study finds that although debt reduction and policy reforms in debtor countries have been important determinants of renewed access to international capi- tal markets, changes in international interest rates have been the dominant factor. We calculate the effects of changes in interniational interest rates for a "typical" debtor country. We conclude that increases in interest rates associated wvith a busi- ness cycle upturn in industrial countries couild depress the secondary market prices of existing debt to levels inconsistent with continlued/ capital inflows. The turnaround in the external financial position of many debtor countries since 1989 has been phenomenal. Improvement is particularly impressive in countries that had completed Brady Plan restructuring of their external debt at the time this article was prepared (Argentina, Costa Rica, Mexico, Nigeria, the Philip- pines, Uruguay, and Venezuela). In the first quarter of 1989 the external debt of these countries sold for an average price of only forty cents on the dollar and private capital inflows were largely restricted to concerted lending or interest arrears. Various plans for dealing with the debt overhang, including the Brady Plan announced on March 10, 1989, were widely characterized as inadequate to restore access to international capital markets. Some observers, in fact, pre- dicted that debtor countries might not return to private international capital markets for a generation (see U.S. Senate 1990). Today the recovery in real economic activity and capital formation in debtor countries is just beginning, but a financial recovery is well under way. These countries have experienced very large private capital inflows, real exchange rate appreciation, stock market booms, and dramatic increases in the prices of their external debt (Calvo, Leiderman, and Reinhart 1993). In some cases capital inflows have been associated with a return to resource transfers to these coun- Michael Dooley and Kenneth Kletzer are with the Department of Economics at the University of California, Santa Cruz; Eduardo Fernandez-Arias is with the Office of the Chief Economist at the Inter- American Development Bank, on leave from the World Bank. Fhe authors acknowledge Nlandu Mamingi for his econometric work, and thank Stijn Claessens and Paul Armington for helpful comments. 01996 The International Bank for Reconstruction and Development/THE WORI D BANK 27 28 I HIV WORLLD BANK I (A) NotMI( RF\'IFW, VOL. 10, No. I tries similar to those recorded in the 1970s, as measured by the emergence of sizable balance of trade deficits. Do we understand enough about the 1982 crisis to predict that renewed accu- mulation of external debt can avoid a repeat of 1982 and the considerable costs that followed for debtor countries? Unless the memories of investors and resi- dents of debtor countries are very short, they must believe that there is a differ- ence in the expected outcome of this new round of international investment. Is the debt crisis dead, as suggested by several observers recently, or is it only sleeping? To understand this turnaround in market access, it is necessary to identify the main factors that can account for the remarkable improvements in debtor coun- tries' creditworthiness. We first argue that secondary market prices for syndi- cated credits are a useful measure of market access. We then show that changes in international interest rates and induced changes in real exchange rates and real domestic interest rates in debtor countries can account for all of the im- provement in secondary market prices after the first round of Brady Plan re- structuring agreements in early 1990. The empirical relationship between sec- ondary market prices and international interest rates is robust to changes in model specification and to the period considered. In particular, the dominance of international interest rates holds both before 1989, when yields on sovereign credits rose over time-and prices declined-and after 1989, when yields gener- ally declined-and prices rose. This is further confirmed by recent developments outside the period of our estimations. For example, during February and March 1994-after this article was prepared-U.S. interest rates increased substan- tially while secondarv market prices dramatically declined.' The decline in real long-term interest rates on dollar-denominated debt is certainly reversible and, in fact, it might very well be reversed in the next year or so. If domestic real interest rates in debtor countries were also to rise, and real exchange rates were to decline-as would be normal-many developing coun- tries would probably again experience debt servicing difficulties. Section I develops the relationship between capital inflows and secondary market prices. Section 11 discusses factors that might explain secondary market prices for developing-country debt. Section III estimates a simple model of sec- ondary market prices. Section IV applies these estimates to a composite Brady Plan country in order to evaluate the source of recent capital inflows. Section V summarizes the results. I. CAPITAL INFLOWS AND SECONDARY MARKET PRICES The secondary market price of commercial bank debt is a useful barometer for country creditworthiness. The secondary market price indicates the climate 1. The ten-year LtIS. Treasury bond rate increased hs 1 5 percenit, from 5.? to 6.6 percent per year, while the marker price index fell by a similar proportion. Dooley. Fernindez-Arias, and Kletzer 29 for private capital inflows to a debtor country, because it reflects both private investors' expectations concerning the ability of debtor governments to service existing debts and yields on alternative international investments. The second- ary market price is a sensitive indicator because it is established in an active market for a relatively homogeneous financial instrument. Furthermore, it is also more up to date and accurate than private capital flow data, useful proper- ties for an indicator. Rising secondary market prices (falling yields) suggest that residents of the debtor country can issue new debt or equity on better terms than those on past debt. The rising prices reflect improvements in country creditworthiness that, to some extent, apply to all forms of external financing. We argue that the im- provement in creditworthiness also results from the worsening in alternative returns in industrial countries. These improvements may fail to be powerful enough to eliminate substantial secondary market discounts and allow countries to regain access to similar commercial bank loans, but may be strong enough to allow access to and better terms for alternative forms of external finance that the market perceives as a safer instrument (for a formal model, see Fernandez- Arias 1995). The important implication for understanding recent capital inflows is that a larger volume of new borrowing, or sales of equity, can credibly be serviced at lower yields. Thus, an improvement in secondary market prices was a precondition for recently observed private capital inflows to debtor countries. Moreover, a return of secondary markets to levels reached in 1989 would cer- tainly stop and probably reverse recent capital inflows. There are two fundamental reasons for changes in the terms on which inves- tors hold new and existing claims on residents of developing countries. The first is changes in yields available on alternative investments as measured here by an appropriate risk-free dollar interest rate. The second is a change in investors' evaluation of the credit risk peculiar to the developing country. While easily observed secondary market prices undoubtedly reflect other factors such as the relative status of government and private debt, our working hypothesis is that the value of sovereign debt is closely related to the investors' overall assessment of the outlook for expected returns on existing and new investments in the debtor country relative to expected returns on alternative investments (Dooley and others 1 990). Capital inflows adjust to equalize alternative returns by financing marginal projects with lower domestic returns (flow adjustment) and by increasing over- all exposure (stock adjustment) (Fernandez-Arias 1995). Although several re- cent papers have attempted to directly explain private capital inflows, this has proven to be a difficult task. The fact that private inflows have been offset by official outflows, generally in the form of increases in international reserve as- sets, makes the existence of a stable relationship between expected yields and private capital flows unlikely. Different policy reactions over time clearly con- taminate reduced-form relationships between expected yields and private capi- tal flows. For this reason we focus on the expected yield of existing commercial 30 THE WORLD BANK ECONOMIC REVIEW. VOL. I0, NO. I bank debt as the best proxy for the terms on which residents of emerging mar- kets can issue new debt and equity. Analysis of this expected yield allows us to trace the underlying determinants of new capital inflows. II. QUANTITATIVE EVALUATION OF FACTORS AFFECTING SECONDARY MARKET PRICES Although there have been a number of recent commentaries on the end of the debt crisis, relatively few quantitative analyses of what went right have been undertaken. While it is natural for disasters to get more attention than for fair weather, a careful evaluation of recent developments can help in analyzing the durability of the improvement. In this section we discuss measures of five fac- tors that might explain secondary market prices for developing-country debt: debt reduction, economic policy reform, international interest rates, domestic interest rates, and exchange rates. Debt Reduction A simple model for secondary market prices sets the market price equal to the ratio of expected present value of debt service payments to the contractual value Figure 1. Debt Prices in Selected Countries, 1986-93 Secondary market price (cents on the dollar) 80 - 70- 60- 50 - 40 30- 20 10 0 - 1987 1988 1989 1990 1991 1992 1993 AVote. Calculated as a weighted average based on commercial bank debt. Argentina, Costa Rica, Mexico, Nigeria, the Philippines, Uruguay, and Venezuela are Brady countries; Albania, Algeria, Angola, Bolivia, Brazil, Bulgaria, Cameroon, Chile, Congo. C6te d'lvoire, Ecuador, Honduras, Hungary, Jamaica, Jordan, Morocco, Nicaragua, Panama, Peru, Poland, Senegal, and the Syrian Arab Republic are non- Brady countries. Source Salomon Brothers (various issues): Internatronal Financial Remieu. (various issues); American Banker, Inc. (various issues); World Bank data. Doolev, Fcrn,indez-Arias, and Kletzer 31 of outstanding debt. It follows that secondary market prices rise when the nu- merator-that is, expected payments-rises relative to the denominator, the con- tractual value. Consequently, these prices are linked to country creditworthi- ness.2 Debt or debt service reduction is expected to increase the price if the present value of expected payments does not fall proportionately with the re- duction in the contractual value of the debt (see Dooley 1988 for a general discussion of buybacks and market prices). As documented by Bacha (1991) and World Bank (I 992), increases in debt prices since the announcement of the Brady Plan in 1989 have beeni larger for Brady Plan countries than for other debtor countries (figure 1). There is much less agreement concerning the quantitative importance of debt reduction. The initial skepticism about the Brady Plan on the part of many aca- demic economists was based on a simple argument. The Plan was voluntary in the sense that banks would not be forced to exchange their existing claims for new claims with a lower expected market value. The implication was that debt reduction would be expensive in the sense that private debt retired by the Brady Plan would be purchased at a price higher than the market price that would prevail if the Plan were not implemented. As pointed out in Dooley (1988), the higlher price would reflect the market value of debt remaining after the agree- ment was implemented if banks could free-ride as in open-market buybacks (see also Bulow and Rogoff 1988). If the banks were in a less strong bargaining position, the price would be lower, and more debt reduction would be possible for the same amount of resources. But the amount of debt reduction would always be limited by the banks' voluntary participation constraint. Given this constraint it is possible to calculate the range of debt reduction that would re- stilt, given the resources available to support the deal. As shown by Claessens, Diwan, and Fernandez-Arias (1992), Brady operations led prices and debt re- duction to fall within the theoretical ranges. Table 1 provides a summary of the debt reduction obtained by various coun- tries to date." The first column shows debt retired as measured by the reduction in the present value of debt service obligations. Debt reduction reflects reduc- tions of contractual debt and interest service as well as collateralization and new money promises. An example is agreements where below-market interest rates 2. The relationship between prices and country risk is distorted in some of the instruments used in Brady operations byv two factors. First, in the numnerator, collateral enhancements increase rhe value of those instruments. Second, in the denominator, below-market interest rates such as those in Brady par hoiids amount to a lower effective contractual value. These two biases counteract each other and may conceivably offset each other in the case of par bonds, which wvould justifv the usual practice of usinig par biind prices. In general, hiowever, thLe two l)iases do niot offset each other and need to be adjusted to obtairn the so-called stripped prices, whose level would better reflect country creditworthiness (see table I for details). The changes in these stripped prices caused by changes in international interest rates canl be approximated by the changes in the prices of Brady bonds because mtost collateral enhancements are also interest-senisitive, and therefore strippinig is not necessary for the econonmetric exercise. 3. After this article was prepared, Brazil, Bulgaria, the Dominican Republic, Ecuador, Jordan, and Poland completed Brady operations in 1994 and 1 99i. Table 1. Debt and Debt Service Reductioni in Brady Plan Counltries, 1990-93 Commnnrcial bank Ov erall net debt-reduction equiValent Total net Additional debt-reduction equivalent Debt prices Debt retired Percentage of pa7ment to banks official loans Millions Percentage Pre-Bradv' Postoperation (millions of commercial (millionzs of (millinns of of U . S. of (cents to a stripped Contry d'ollars) bank debt T.S. dollars) U.S. diollars) dollars' total debt dollar) (cents to a dollar) Argentina 10,723 37 3,732 2,117 8,606 14 18 63 Costa Rica 1,166 73 225 177 989 21 12 39 Mexico 19,033 40 6,812 3,732 15,301 16 .36 51 Nigeria 4,221 79 1,681 0 4,221 14 21 45 Plhilippinies' 3.553 54 1,45 1 1 54 3,399 1 2 40 76 Uruguay 807 50 413 140 667 1S 56 73 Venezuela 5,153 27 1.949 687 4.466 14 37 59 Total 44,656 40 16,263 ,007 .37,649 15 .31 57 a. The amount is obtained by subtracting the additional official loans from the debt retired. b. The debt reductioni was comiipleted in two phases. c. IPrice in the monith before the Bradv Plall was annoLInced (March 10, 1989). Source: Claessens, Diwan, and Fernandez-Arias ( 1992) and authors' calculations. Doolev. Ferndndez-Arias, and Kletzer 33 on collateralized par Brady bonds were exchanged for old floating-rate debt. We calculate the difference in the present value of the debt service obligations of the two bonds on the exchange day, assuming that each would be serviced in full as contracted. The methodology used is almost identical to the one used in Claessens, Diwan, and Fernandez-Arias (1992). The only difference is the treat- ment of additional new money, whose negative effect on debt reduction is esti- mated as a fraction of its nominal value (the fraction being the ex ante price). Because banks' promises of new money were often conditional on countries' serving interest over a period of time (not a sure thing in the absence of the deal, as reflected in low prices), this estimation is probably better. The third column in table 1 shows the net payment received by commercial banks. This cash was used to purchase collateral for new bonds or more directly for buybacks. In general, however, the reduction in the contractual present value of debt was largely independent of the financial engineering involved. These calculations indicate that the amount of resources devoted to the agreements are more than a third of the reduction in the contracted present value of private debt. Substantial additional official lending partially offset the reduction in com- mercial bank debt. The fourth column shows the dollar amount of additional loans made to the debtor government by international organizations and credi- tor governments to support the Brady Plan. Thus the niet debt reduction repre- sents only 15 percent of total debt (fifth and sixth columns). It is not difficult to see why many analysts doubted that this level of debt reduction would be deci- sive in reestablishing access to capital markets. One way to evaluate the direct effects of debt and debt service reduction on secondary market prices is to analyze the market price of debt remaining after the restructuring. Prices of instruments are distorted by various features and attachments, such as collateral, new money promises, and below-market inter- est rates. Therefore, the last column in table I shows stripped prices, that is, the prices right after the operation, adjusted for these distortions. These prices are a good indication of the market view on country creditworthiness once the ben- efits of the operation are fully factored in. (Like the calculation of debt reduc- tion equivalent, the methodology for estimatilg stripped prices is taken from Claessens, Diwan, and Fernandez-Arias 1992, except for estimating the impact of additional new money.) If future repayments to commercial banks are positively linked to the country's future performance, then the efficiency gains of these debt and debt service reduction operations can be gauged by analyzing the impact of the operation on prices. In fact, in the absence of efficiency gains, in propor- tional terms (stripped) prices would not be expected to increase beyond the decrease in commercial bank debt (second column in table 1). As pointed out by Dooley and others (1990), a full evaluation of the impact of debt reduc- tion on the value of remaining private debt should consider the relative se- niority of the various types of debt and the probability that the debtor would have received the loans for another purpose. A hypothesis consistent with 34 THE WORLI) BANK F(ONOMIC RFVIFEW. VOL. IiR, No. I the findings in Demirguc-Kunt and Fernandez-Arias (1992) and in Bulow, Rogoff, and Bevilaqua (1992) is that all creditors have the same implicit seniority and share the net present value of repayments in proportion to exposure. If this is true, then, in the absence of efficiency gains, (stripped) prices would not increase beyond the decrease in total debt (sixth column in table 1). Any excess price increase over the no-efficiency-gain benchmark could then be safely attributed to efficiency gains. Unfortunately, the task of estimating the increase in prices caused by the debt-reduction operation is extremely difficult because the appropriate counterfactual price-the price prevailing in the absence of the operation-is not observable. Long before the operation was consummated, prices reflected the market expectations on the outcome of the future operation, and thereby contaminated the observed prices to an unknown extent. For example, if the last price quoted before the operation incorporates a perfect forecast of the opera- tion, its comparison with the stripped price does not provide any meaningful information on the effects of a given operation. Prices before the Bradv announcement in March 1989 may not be subject to this contamination, but they do not reflect the changes in economic fundamen- tals over the period leading to the operation date. For this reason, results based on these prices (shown in the seventh column of table 1) need to be taken with caution. Nevertheless, as analyzed in the next section, the evidence shows sig- nificant variation only in international interest rates after most of the first-round Brady operations had taken place. Therefore, except for the recent operations and especially in Argentina, estimations and inferences made on the basis of prices prevailing before the Brady announcemenit appear reasonable. Economnic Policy Reform It is plausible that the conditionality associated with the Brady Plan agree- ments explains the increased market value of existing debt and the turnaround in access to external markets. It is difficult to quantifv the effects of economic reform on market valuations of external debt, but it certainly appears that poli- cies changed for the better in Brady Plan countries. The widespread adoption of market-oriented reform programs along with aggressive fiscal reform may have been an additional important channel through which the Brady Plan workouts improved the financial position of debtor countries. It is perhaps not surprising that creditor governments emphasized this aspect of the plan. What may have been surprising, however, was how consistently and aggressively some of the debtor countries implemented fiscal reform changes. This suggests that the im- pact of fiscal reforms was not fully credible at the time of the debt exchanges (actual execution of the Brady operation). The effect of improved fiscal policies on secondary market prices may have been gradually incorporated into market prices in countries where the reform in fact occurred. One measure of a number of important policy changes is the increase in gov- ernment revenue net of expenditures other than debt service-what is usually Doolev. Ferndndez-Arias, and Kletzer 35 Table 2. The Primary, Fiscal Surplus (PFS) and the Operational Fiscal Surplus (OFS) in Selected Countries, 1985-92 (percentage of GDP) Country 19S 5 1986 1987 1988 1989 1990 1991 1992 Argentina Primary fiscal surplus 0.8 1.8 -0.9 -1.0 -6.3 1.6 3.5 3.8 Operational fiscal surplus -6.0 -4.7 -S.6 -6.3 -21.9 -2.9 -0.2 1.5 Brazil Primary fiscal surplus 2.1 0.6 -2.8 -0.5 -0.5 2.2 1.0 2.5 Operational fiscal surplus -4.3 -3.6 -5.7 -4.8 -6.9 1.3 -2.2 -2.2 Chile Primary fiscal surplis 0.6 0.5 251 6.6 7.5 5.0 2.2 - Operational fiscal surplus -2.9 -6.0 -I1.0 -1.2 3.1 1.0 -1.2 - Mexico Primary fiscal surpitis 3.9 2.2 5.8 8.1 8.4 7.6 8.8 8.7 Operational fiscal surplUs -3.i3 7.0 1.8 -.3.6 -1.7 2.3 6.7 6.0 Morocco Primary fiscal surplus 0.7 1.2 1.2 2.2 1.8 5.6 5.0 - Operational fiscal surplus -5. 5 -6.8 -2.6 -1.1 -1.4 2.1 1.5 - Nigeria Primary fiscal surplus 4.7 2.6 2.4 -0.1 5.5 6.5 5.8 - Operational fiscal sLUrplus -1.8 -5.4 -4.3 -5.9 0.5 0.4 0.2 - Philippines Primilary fiscal surplus 2.4 -1.3 2.7 3.2 1.4 1.1 1.0 - Operational fiscal surplus -2.1 -6.1 -0 --0.1 -1.6 -2.7 -3.0 - Venezuela Primary fiscal surplus 4.7 0.9 -1.1 -6.1 3.8 6.1 7.1 -0.5 Operational fiscal surplus -0.2 -9.9 -3.6 -9.9 -1.0 2.1 3.5 -4.5 - Not available. Source: Goldman Sachs (I1991, 1992). called the primary fiscal surplus.4 As shown in table 2, some Brady Plan debtors have made very impressive budgetary progress and can finance a considerable percentage of debt service payments through taxation rather than through addi- tional borrowing. Another useful measure of fiscal performance is the opera- tional fiscal surplus (OFS). This is the primary surplus less real interest payments on both domestic and external debt. Improvements in this surplus relative to the primary surplus are caused by a fall in domestic or international real interest rates or a fall in the stock of debt. The impressive improvement of the opera- tional balances in table 2 reflects the combined impact of all of these factors. In 4. Proceeds from privatization are included as reveuille. 36 THF WORLD BANK ECONOMIC REVIEW, VOL. 10. NO. I Table 3. The Public Debt Ratio in Selected Countries, 1985-92 (percentage of GDP) Country 1985 1986 1987 1988 1989 1990 1991 1992 Argentina 72.2 78.6 89.9 95.9 112.3 94.3 68.5 62.0 Brazil 50.6 48.0 48.7 45.6 42.2 40.1 47.1 46.5 Chile 76.9 85.6 83.7 67.7 52.2 39.8 - - Hungary 42.5 46.7 56.9 52.6 55.5 54.0 - - Mexico 51.9 59.2 54.5 61.7 56.1 48.5 35.0 25.0 Morocco 137.1 127.5 136.3 125.3 117.3 105.2 - - Nigeria 50.1 88.2 133.6 118.7 113.3 114.3 - - Philippines 57.6 69.5 76.6 73.5 67.3 71.7 - - Poland 43.2 48.5 67.2 65.0 73.0 88.9 - - Venezuela 41.2 59.9 54.6 53.8 70.4 54.1 46.3 52.0 - Not available. Note: The public debt ratio incluides domestic arid external indebtedness of the public sector minus official reserves. Source: Goldmaii Sachs (1991, 19921. the empirical work that follows we assume that changes in these fiscal balances are correlated with a variety of policy reforms that are difficult to quantify. Although these measures are quite incomplete, it seems unlikely that strong changes for the better in policy regimes would not be closely related to improve- ments in these fiscal balances. Improvement in the operational balance has been much more pronounced compared with improvement in the primary balance. An important challenge for evaluating the future is to identify what part of the reduction in real interest payments is a permanent part of the debtor countries' positions. One aspect that is clearly permanent is real debt amortization. A surplus for this operational budget balance for past years means that the real value of outstanding debt is being reduced. As shown in table 3, some debtor countries have made substan- tial gains in reducing the real value of their net government debt through a normal amortization of domestic and foreign debt. For some countries this has been more important than the debt reduction discussed above. It is also possible that the fiscal reforn will generate a permanent reduction in the default premia incorporated in interest rates. International Interest Rates Another potential source of improvement in debtor countries' positions has been the change in the external environment. The dominant change after 1989 was a fall in nominal and real interest rates in the United States and, to a certain extent, in other major industrial countries (figure 2). As shown in Dooley and Stone (1993), the rise in international interest rates is the only variable in a regression analysis that has much power in explaining the widespread decline in secondary market prices through 1989. This result is consistent with the hy- pothesis that the expected present value of payments by debtor countries fell as Doolev. Ferntindez-Arzas, and Kletzer 37 Figure 2. The Government Bond Yield and the Real Interest Rate in Gernany, Japan. aindl the United States, 19865-93 Goivernment Bond Yield Real Interest Ratel Percent Percent per year Germany per year Germany 9 7 g ~~~~~~~~~~~~~~~~~~~~~6 9 1986 87 88 89 90 91 92 93 1986 87 88 89 90 91 92 93 Percent Plercent pet year Japan per year Japan 96 8 9 6 - 1986 87 88 89 90 91 92 93 1986 87 88 89 90 91 92 93 Percent Percent per year UTnited States per year lJnited States 6 8 O -, ~~~~~~~~~~~~~~~3 2 6 0 1986 87 88 89 90 91 92 93 1986 87 88 89 90 91 92 93 a. Line 61 in IMrI: (various years). b. Six-month London interbank offered rate (LIBR10 (line 6Oeb) minus inflation (calculated as the change in the consumer price iniclex, line 64). Source IMNI (varlIIus years) international interest rates rose. Thus, the rise in debt prices after the Brady speech would be consistent with the observed fall in market interest rates. The potential importance of the fall in international interest rates arises from two sources. First, there are good theoretical reasons to believe that the value of both fixed- and floating-rate Brady bonds should rise more than proportionally to percentage declines in international interest rates. Second, domestic interest 38 IHF WORI RBANK F(Ot)NONIl RFVIPW, Vol 10, No. I rates paid by debtor governments should fall with international rates, and there are good theoretical reasons to predict that the reduction will be more than proportional. To evaluate the effects of changes in international interest rates, it is necessary to identify a relevant 'discount rate" at which investors translate expected pay- ments from the debtor government into a present value. Since most of the exter- nal debt is denominated in U.S. dollars, the appropriate discount rate is a risk- free real interest rate available on1 a dollar-denominated investment that is similar in terms of maturity and in the terms on which the contractual interest rate is adlusted over time. This is not a straightforward problem. In particular, it would at first seem natural to compare floating-rate sovereigni debt to floating-rate risk- free debt. The problem with this approach is that, for risk-free floating-rate in- strumenits, changes in market interest rates alter the nominal value of expected payments in future time periods. But this is exactly offset by the change in the discount rate so that the present value of these payments is unchanged. With floating-rate sovereign credits that trade at a considerable discount, the effect of a change in the real risk-free rate is quite different. Assuming that the change in the real risk-free rate does not change the government's ability or willingness to puay, the value of expected paymiients in future time periods does not change. It follows that the present value of these payments does change. Thus, both floating-rate sovereign credits and stripped prices respond to a fall in international interest rates in a manner usually associated with fixed-rate long- term bonds. To the extent that repayments are shared by foreign creditors in proportion to contractual debt service, the response of fixed-rate sovereign credits would be even more pronounced because it would increase the share of fixed- rate debt service oblig,ations in total debt service. If future payments are expected to grow over time, as can be expected in a growing economy, then the increase in their present value would be proportion- ally larger than the decrease in the risk-free rate. Furthermore, if the foreign debt is lower in priority of payment to other types of debtor government expen- diture, secondary market prices will tend to rise by more than the percentage increase in the present value of total expected payments. This is a potentially important aspect in unlderstanditng the relationship between international inter- est rates and secondary market prices. Unlike substantially risk-free instruments, a fall in the discount rate increases the present value of both floating- and fixed- rate debt of overindebted counltries. Domnestic Interest Rates Recent empirical research has documented a strong link between interna- tional interest rates and domestic rates in developing countries (Frankel 1994; Glick and Moreno 1994). Most internal debt is rolled over several times a year in debtor countries, and so real debt service payments are very sensitive to changes in domestic real interest rates. This is an interesting part of these governments' expenditures because, relative to real interest payments to foreign creditors, real Doolei. Fern,indez-Arias, and Kletzer 3 9 interest rates paid on their domestic debt show a much higher variance and much higher average levels before 1990. Although internal debt is typically smaller than external debt for these countries, changes in ex post real domestic interest payments have been an important component of total debt service costs. A rise in domestic debt service payments should, for a given overall capacity to pay, reduce expected payments on external debt and in turn lower secondary market prices for external debt. If changes in international interest rates gener- ate qualitatively similar changes in domestic rates, as would be expected if capi- tal markets are at all integrated, this would clearly reinforce the effect of inter- national interest rates on secondary market prices. Exchange Rates Government revenue in domestic currency can cover greater debt service pay- ments if the foreign currency value of revenues rises, as happens wheni the local currency appreciates. Other things being equal, the real appreciation of curren- cies in debtor countries, shown in figure 3, increased the dollar value of govern- ment revenues devoted to external debt service. As with the other variables dis- cussed above, the relevant measure of the real exchange rate is that expected to prevail over the life of the contract. For lack of a better prediction, we can take the current value as an unbiased, but certainly poor, prediction of its future values. Of course, the real exchange rate is not an exogenous variable, so other things are probably not equal. Our assumption that the real exchange rate fol- lows a random walk is a weak but reasonable one, because structural models of exchange rate determination have not performed better than the random walk. It is also possible that changes in the real exchange rate do affect the domestic currency value of the fiscal deficit. For example, the dollar value of oil revenues does not change following a real exchange rate shock. In the empirical work we simply expect a positive relation between the terms of trade and debt prices. As with domestic interest rates, it is also important to consider the relation- ship between exchange rates and internatiotnal interest rates. If capital inflows associated with low internationial interest rates induce exchange rate apprecia- tion, it follows that we underestimate the effects of exogenous changes in inter- national interest rates on debt prices. Thus, the assumption that real exchange rates are unrelated to other variables in the model probably works against our main hypothesis. 111. A SIMPLE MODEI. OF PRI(.F: CHANGES The arguments developed in section 11 suggest the following regression hy- pothesis: (1) p,t c + a,+ J3LTX,t+ yLTG,t + 8RCTit + Er, + u,, where t = 1, 2, . T is the time index; i = 1, 2, . . ., N is the country index; p is the logarithm of secondary market prices; LTX is the logarithm of the ratio of 40 IHE WORLD BANK E(:O)NOMiC REVIEW, VOL. 10, NO. I Figure 3. Index of Real EfIective Exchange Rates in Selected Debtor Countries, 1986-93 Argentina Bolivia 85 5 75 80 65 75 55 ~~~~~~~~~~~70 45 1~~~~~~~~~~~6 35 6 25 60 1986 87 88 89 90 91 92 93 1986 87 88 89 90 91 92 93 Brazil Clhile 180- ~~~~~~~~~~63- i6o- ~~~~~~~~~~~59 55- 120-~~~~~~~~~~~~5 Colombia Ecuador 72- o 67- 5 70- 62- 65- 57- 60- 52 ~~~~~~~~~~~~50 47 4 1986 87 88 89 90 91 92 93 1986 87 88 89 90 91 92 93 Mexico Peru 100 - 33 -h 60l - 130- 50 80 1986 87 88 89 90 91 92 93 1986 8-7 88 89 90 91 92 93 1 Iruguay Venezuela 83 -10 78-90 73- 0 68- 0 63- 0 583 0 1986 87 88 89 90 91 92 93 1986 87 88 89 90 91 92 93 Note. An increase in tle index denotes a real exchange rate appreciation. Source. IMF Information Notice Svstem data base. )ovetv, F-ernindez-Arias, and Kletzer 41 total long-term debt to exports; LTG is the logarithm of the ratio of total long- term debt to GNP; RCT is the logarithm of the ratio of commercial debt to total long-term debt; r is the logarithm of the long-runi (ten-year) U.S. interest rate; c is the common constant term; c(., are the counltry specific intercept terms; and u is the usual error term. Since our purpose is to show that the substantial increase in the debt prices of Brady countries after 1989 can be easily explained by changes in international interest rates (and the purely arithmetic effect of debt reduction), and therefore does not point to fundamental improvements in the countries' economic pros- pects, we chose to use a parsimonious empirical model. The above specification simply adds the international interest rate to the usual basic determinianits of commercial bank debt prices in the literature: the most common indicators of country creditworthiness (the debt-to-exports and debt-to-GNP ratios) and the share of commercial bank debt in total debt. In line with the empirical literature, the explanatory variables were assumed to be statistically exogenous. This assumption is clearly justified in the case of the international interest rate, the key variable introduced in this analysis. The use of the change in the debt share caused hy exchange rate effects as an instru- ment for the commercial bank debt share indicates that that explanatory vari- able, which is largely predetermined, can be also assumed to be exogenous (see Bulow, Rogoff, and Bevilaqua 1992). The resource variables, exports and (NP, are flows during the year prior to the point in time at wlich prices and debt stocks are measured, and are therefore predetermined. Bulow, Rogoff, and Bevilaqua (1992) also experiment with instruments for these and other vari- ables and conclude that instrumenital variable estimation is not needed. The underlying notion behinid these models is that, under credit rationing, the value of commercial bank claims as a wlhole amounts to a piece of the country's re- sources, which are in turn largely exogenous (there is an analogy here witlh the value of claims against an insolvent firmni. The above specification in terms of price rather tllani market value (price times stock) is best seen as the rescaling of an underlying value equation corrected for stock-related heteroskedasticitv. As a result, the equation was estimated based on a least-squares method, taking into consideration the paniel nature of the observations (see the appendix for more details).' As shown in Dooley and Stone (1993), conventional regressors for seconidary market prices, such as the debt-export and debt-c(DP ratios, measures of the composition of debt, and fiscal variables, explain cross-section differences in prices from 1986 througIh 1990. But the international interest rate is the domi- 5. Additional variables halve beein used in the empirical literature to explain secondary debt prices. sulch as the reserve-import ratio or the proportion of debt n iarrears (see, for example, Stone 1991 ). The endogeneity of these varialiles is clearly a significant potential problem. External variables other than international interest rates may be also relevant for explaining the co[inovemenit of prices across counitries. To the extent that externial variables are niot correlated with interest rates, their exclusion should not indUce seriiOtis estiztimoti biases. 42 THF WORLI) BANK E(CONOMIC REVIEW, VOL. 10, No. I Table 4. The Impact of Debt and Interest Rate Variables on Secondary Market Prices in Twenty Developing Countries, 1986-92 Variable Coefficient Ratio of total long-term debt to exports, LTX -0.50 (-3.10) Ratio of total long-term debt to (.NP, LTG -0.36 (-2.30) Ratio of commercial debt to rotal long-term debt, RCT 0.09 (0.86) Long-run (ten-year) LJ.S. interest rate, r -0.87 (-3.17) Constant term, c 5.67 (8.70) R' 0.36 Adiusted R2 0.22 Note: Generalized least squares ((tLs! was used to estimate the panel regression equation. The dependent variable is secondary marker prices. All variables are in logarithms. t-ratios are in parentheses. The twenty countries in the sample are Argentina, Bolivia, Brazil, Chile, Colombia, C6te d'lvoire, Costa Rica, Ecuador, Guatemala, Jamaica, Mexico, Morocco, Nicaragua, Nigeria, Panama, Peru, the Philippines, Senegal, Uruguay, and Veenezuela. Source: Authors' calculations. Table 5. The Impact of Debt, Interest Rate, and Fiscal Surplus Variables on Secondary Market Prices in Seven Developing Countries, 1986-92 Coefficients for model Excludling Including Including includitng Variable PFS and OFS PFS and OFS OFS PFS Ratio of total long-term debt to -0.30 -0.23 -0.24 -0.24 exports, LTX (0.16) (0.19) (0.19) (0.18) Ratio of total long-term debt to GNP, LTG 0.22 0.20 0.22 0.20 (0.14) (0.16) (0.15) (0.15) Ratio of commercial debt to total -0.05 -0.06 -0.07 -0.05 long-term debt, RCT (0.13) (0.14) (0.1 4) (0.13) Long-run (ten-year) U.S. interest rate, r -2.66 -2.57 -2.61 -2.58 (0.47) (0.48) (0.47) (0.48) Transformed primary fiscal surplus, LPFS' -0.03 0.32 (Lt 19) (0.50) Transformed operational fiscal surplus, LOFS' 0.33 0.37 (0.74) (0.80) Note: Generalized least squares ((.LS) was used to estimate the panel regression. The dependent variable is secondary market prices. All variables are in logarithms. Standard deviations are in parentheses. The seven countries in the sample are Argentina, Brazil, Chile, Mexico, Morocco, the Philippines, and Venezuela. a. LPFS and LOFS are positive transformations of PFS and OFS as a fraction of debt outstanding. Source: Authors' calculations. nant determinant of the time-series behavior of prices. A similar finding is re- ported in Cohen and Portes (1990). An important reason to doubt this result is a clear common trend for prices and interest rates over the 1986-89 time pe- riod. In this article we extend the sample period to 1992, a period in which there was a clear reversal in the trend for both interest rates and prices. Results re- Doolcv, Fernanndee-Arias, azd Kletzer 4? Table 6. The Impact of Debt and Interest Rate Variables on1 Secondary Market Prices in Seventeen Developing Couintries, 1986-89 and 1989-92 (full sample) Variable 1986-89 1989-92 Ratio of total long-term dehr to exports. LTX -0.071 -0.48 I--3.76) (-2.47) Ratio of total long-term debt to GNP, LTG -0.0,53 -0.55 (-.3.13) (-3.24) Ratio of commercial debt to total long-term debt, RCT -0.20 -0.(7 (--1.4-7 (-0.55) Long-run (ten-year) U.S. interest rate, r -4.91 -0.75) (--9.25) (-3.24) Constant term, c 1 3.96 4.96 12.0)1 1 (8.(01) R' 0,70 (0.52 Adjusted R` 0.6 1 0.33 Note: Generalized least squares (GLS) wvas used to estimate the panel regression equation. The dependenit variable is secondarv market prices. All variables are in logarithms. t-ritio, are in parentheses. The seventeen countries in the sample are Argentina. Bolivia. Bra/id Chile. CY,re dI'Ivoire, Costa Ric.i, ECu.Idldr, Jamaica, Mexico, Mvlorocco, Nicaragua, Panama, Perul. rhl l'hilippines, Senegal, UIrHig.lvl, ind Vellen/nLcla. Souirce: Au.thors calcul.ations. ported in table 4 summarize panel regressions for annual data for twenty devel- oping countries over the 1986-92 time period. Results for seven countries for which we have data for fiscal balances are reported in table 5. (See the appendix for details on data and econometric methods.) For the larger sample (table 4), the conventional measures of debt relative to the economic resources available to service the debt have the expected signs, and are statistically significant at conventional levels. These variables presum- ably capture the impact of debt reduction and improvements in the debt service capacity of the debtor country. For the smaller sample of countries (table 5) the basic model is less satisfactory; inclusion of the primarv fiscal surplus does not improve the statistical properties of the basic model, and is not a significant variable. This is consistent with results reported in Dooley and Stone (1993). Our primary interest, however, is on the size and stability of the interest rate effect. As shown in table 4, for the larger sample the interest rate has the ex- pected negative sign and is near the expected value of negative unit elasticity. That is, a 1 percent change in the long-term U.S. Treasury bond interest rate, for example from 5 to 5.05 percent, genierates about a I percent fall in market price. To test the robustness of this result, we also divided the larger sample into two periods roughly corresponding to the period of generally falling prices be- fore 1989 and generally rising prices thereafter (table 6). Again, the interest elasticity has the expected sign and is statistically significant, although the abso- lute size of the elasticity in the earlier time period is implausibly large. While the interest elasticity is -4.91 in the 1986-89 period, it is -0.75 for the period 1989-92. This discrepancy can perhaps be explained in terms of an omitted 44 THE WORLD BANK FCoNOMIC REVIEW, VOL. 1(, NO. I variable that would measure increasing investor pessimism. Thus, the elasticity is biased downwards when interest rates made a negative contribution (1986-89) and upwards when interest rates made a positive contribution (1989-92). Such an interpretation is further confirmed when a time dummy is included in the specification of table 4: the time dummy is significantly negative, and the over- all estimated interest rate elasticity becomes -1.70. Interest rates exerted a sub- stantial effect in the expected direction during both periods. This is reassuring because interest rates increased in the first period and declined in the second period. Our interpretation of this evidence is that changes in international interest rates have had an important influence on market prices of existing debt of de- veloping countries and, in turn, on the reentry of residents of these countries to international credit markets. The remarkably parallel evolution of prices in Brady and non-Brady countries shown in figure 1 further confirms the notion that the international interest rates are the key underlying factor. IV. A SIMULATION EXERCISE In this section we use the results reported in the previous section to assess the importance of interest rate changes and other factors to the evolution of second- ary market prices for a composite Brady Plan country. The econometric results support the use of the following simplified model for country i: (2) p,, = (c,*Bi,)/r, where B^,= (x,,) ! (gi,) where x denotes the exports-to-debt ratio and g denotes the GNP-to-debt ratio. This simple model has a unitary interest rate elasticity and is homogeneous in the country-specific variables exports, GNP, and debt. One implication of this model is that what matters for the price of commercial bank debt is total debt rather than commercial bank debt. This is similar to findings in other empirical studies, for example, Bulow, Rogoff, and Bevilaqua (1992). We stop short of concluding that all creditors have equal seniority status, however, because this condition is necessary but not sufficient unless restrictive burden-sharing mod- els are assumed (for a discussion, see Demirguc,-Kunt and Fernandez-Arias 1992). A more concrete assessment of the factors discussed in section II can be gen- erated by this simple model. For the purpose of illustration, the Brady deals concluded in 1990-92 (Costa Rica, Mexico, Nigeria, the Philippines-Phases I and II, Uruguay, and Venezuela) are aggregated, adding up all values as if they were a single country.' Consider this composite Brady country in March 1989, when the broad outline of the plan was presented to the market in a speech delivered by Secretary Brady. The contractual value of the outstanding commer- cial bank debt was about $81 billion, and the average market price was about 6. The Argentina operation is not included because, as noted above, its analysis is complicated by the inapplicability of pre-Brady prices as benchmarks. Although these problems are also present to some extent in other Brady operations, the size of the Argentina operation may significantly distort the average. I)nDlcv, kernindez-Arias. and K o_-_ - 180 -160 7 - 140 6 - 120 S - 100 4- -80 3 - 60 2 - -4 1 --20 o- l l I I 1989 199( 1991 1992 1993 U-S. government boncl (ten-yeat- tield) U USrT-bill (thiee-miionitih yielcl) Long-term private net inflowvsa W Total ptivate net inilowsa a. Includes all developing countries in the Debtor Reporting Systeml oi the World Bank as reportecd in World Bank (1994). Soutrce: IMr (various years) and World Bank (1994). levels of capital inflows across countries conifirm the relevance of couLitry- specific characteristics, but they do not imply that cbaniges in such country- specific factors caused the inflows, as implied by the pull story.4 Despite these arguments supporting the pishi view, the most reasonable con- clusion to draw from existing evidence is that, althoughi decreases in interna- tional interest rates R have undoubtedly been important in explaining the ob- served magnitude of increases in F for many countries, we caninot infer, for several reasons, that changes in domestic factors-or, for that matter, in exter- nal variables other than rates of return on finanicial assets-have not played a role as well. 4. It is important to note that even a situatton in which some countries receive no new capital inflows is consistent with the push view. The solition for F fromii equation I naY entail an extremely low level of capital inflows or capital outflows (negative values of various componienits otf F), implying transfers of resources that the country 1i unwillug to undlertake, tinder such circumstances the solition for F would be stibject to an inequality constrainr of the form F > F1. It this conistrainit is biniding, such voluntary capital flows Would cease, and equationi 1 Would hecome an inequality, nO longei- determinintig an1y observed (involuntarv) capital flows. As long as fltuctuatiois in external conditions leave this conistiaint bindinig. capital inflows would be unchanged. 62 FHE WORLD BANK ECONO)MIC REVIFW, VOL. 1), NO. I The main reason is that such pull variables are hard to measure. In theory, inflows are endogenous with respect to a wide range of domestic policies, and no single indicator is likely to represent the broad thrust of such policies with the same degree of accuracy as external interest rates do for foreign financial conditions. Indeed, pull factors have been proxied in very rough ways in past studies. In Fernindez-Arias (1995), for example, pull factors are proxied by a shifting intercept term. In Dooley, Fernindez-Arias, and Kletzer (1996) their contribution is captured in the unexplained portion of the secondary debt price, a procedure that is sensitive to the validity of the underlying burdeni-sharing model. A second reason is that much of the existing literature has been restricted to explaining portfolio flows. As shown in section 1, foreign direct investment has been at least as important in many cases, and this type of flow may be more sensitive to domestic factors than the more-liquid portfolio flows. Moreover, a complete story about the factors driving the new inflows must account for changes in the composition of assets acquired by external creditors. These changes present a dramatic contrast between the current and previous inflow episodes. The push story based simply on low U.S. interest rates fails to address this issue. External shocks have been proxied by foreign rates of return in the empirical literature. As a result, the role of structural changes in creditor- country financial markets, which have eased access for developing-country bor- rowers, has not been considered. The existing literature is unable to distinguish between changes in the degree of financial integration (except for factors per- taining to country default risk) and changes in relative ex ante rates of return. The distinction is crucial for the central question that has motivated this litera- ture-the question of sustainability. To the extent that the new flows represent a one-time portfolio adjustment driven by permanent changes in the degree of world financial integration, their high level is not sustainable, but they are less likely to be reversed than if they are driven by temporarily low U.S. interest rates. Thus a consistent story about the factors driving and directing the recent surge in capital inflows should feature some combination of push and pull fac- tors. One such story would proceed as follows. The combination of low interest rates and recession forced low rates of return on industrial-country assets (par- ticularly in the United States), creating an incipient capital outflow as investors in these countries sought higher-yielding assets for their portfolios. The restora- tion of perceived creditworthiness was necessary for potential debtor countries to have access to these funds, and thus capital flowed initially to those countries whose creditworthiness was not severely impaired during the 1980s-largely the rapidly growing countries in East Asia that never suffered a debt crisis. The Brady Plan, announced in mid-1989, broadened the geographic scope for such inflows to include the heavily indebted countries in Latin America, in part by writing down the face value of debt, in part by supporting policy adjustments, and in part by providing information externalities, leading to bandwagon ef- fects. Where none of these factors have come into play-that is, in most of Sub- Saharan Africa-capital inflows have not materialized. -enrmindez -Arias and Montiel 63 Implications for Policv Although the weighing of push and pull factors is informative for policy, it represents at best a point of departure for policy analysis because the mapping from pull or push views to policy is highly imperfect. As indicated above, policy design requires the specific identification of both causal factors and country circumstances. The implicit assumption that capital inflows attracted by im- proved domestic policies do not present a policy problem, but those driven by expansionary monetary policy abroad do, is unwarranted. Even a pull exerted by moving from a distorted to a completely undistorted domestic microeconomic environment could generate macroeconomic instability, calling for a macroeconomic policy response. On the other hand, a pull generated by either a partial removal of domestic distortions or the introduction of new distortions could be welfare reducing on microeconomic grounds as well. Similarly, the implications for policy of an inflow generated bv a foreign push are ambiguous in general, depending crucially on the characteristics of the domestic economy. IV. SUSTAINABILITM The concern that inflows may threaten macroeconomic stability arises in part from a fear that the flows may be transitory. Although even permanent inflows can create adjustment problems, inflows that are not sustained can potentially destabilize the domestic economy when they arrive and wheni they depart. The issue of sustainability can be decomposed into two parts. First, what is the ex- pected time path of the factors driving the inflow episode (for example, how long are the conditions likely to persist)? Second, what are the corresponding implications for capital inflows? Specifically, is the alternative to the current level of inflows a continuation of inflows at a reduced rate (soft landing), a cessation of inflows (hard landing), or pressure for the reversal of capital flows and a balance of payments crisis (crash)? Unfortunlately, the literature to date has shed little light on these questions, apart from the identification of causes. In this section we address the issue in a preliminary way. The first of the two questions is of interest to policymakers in the recipient coun- trv to the extent that the factors driving inflows are exogenous to their actions. As indicated in the previous section, evidence suggests that a substantial external shock in the form of lower interest rates in the lUnited States has been a key driving factor determining the magnitude of capital flows to creditworthy developing countries. Empirically, therefore, the current inflow episode contains an important exogenous component. This being the case, it is meaninigful to ask how long the favorable external shock is expected to last and what the likelv consequences would be of a reversal of these external circumstances or of domestic policies. Dzurationt of the External Shock One way to gauge the likely duration of the foreign interest rate shock is by examining the implicit predictions of future interest rates captured in the term 64 THE WORLI) BANK FCONOM]( RFVIF'I+, VOL. 10, No. I structure. Interest rates steadily declined in the period 1989-93 and started to increase in 1994. As of the third quarter of 1994, when this article was pre- pared, the term structure of interest rates for the United States suggested that interest rates were expected to rise during the subsequent five years, approach- ing their 1989 levels. Thus markets did not expect the favorable external inter- est rate shock to persist. Increases in interest rates in creditor countries would, of course, reduce the incentives for reallocating portfolios to developing countries. Equations 1 and 3 suggest that such incentives would be reduced through increases in the opportu- nity cost of funds and increases in country risk. Thus, both mechanisms have a bearing on the sustainability of inflows. Consider first country risk, which has been the key to extreme forms of unsustainability, such as the debt crisis. Equation 3 shows that this mechanism operates through the market valuation of the present and future resources avail- able to the country to service its external liabilities. Beyond a threshold point, country risk may be too high to sustain voluntary inflows. In this case equation 1 would yield inflow levels less than what the domestic economy could feasibly generate. If so, capital rationing and financial crisis are the likely consequences. Below we construct a simple creditworthiness index to measure the pressure on repayment capacity exerted by the service of foreign liabilities, which can be used to shed light on the likelihood of a crisis. Ani Index of Creditwortbiness Because in the current inflow episode foreign liabilities have primarily been incurred by the private sector (see table 3) and to a large extent denominated in domestic currency, country risk is likely to be associated with balance of pay- ments crises, the attendant likelihood of devaluation, and the imposition of capital controls rather than with fiscal problems. This was illustrated by the recent Mexican crisis. (For the role of fiscal problems in the previous inflow episode, see Montiel 1993.) Under these circumstances the country's repayment capacity can be taken to depend on its ability to generate a trade surplus-that is, to expand exports and contract imports-which depends on its potential to pro- duce traded goods. From the perspective of external creditors, the operational significance of the quality of the domestic policy environment is reflected in this variable. Because the present and future values of maximum trade surpluses are unobservable, for the purpose of constructing a sustainability index, capacity to pay can be proxied by a fraction f of total production of traded goods, T. The present value of this capacity to pay can be compared with an accumu- lated stock of foreign liabilities S to assess whether the country's resources can support the accumulation of additional liabilities. Such a comparison forms the basis for our operational measure of creditworthiness. The present value of re- sources is given by an expression similar to equation 3 with Y equaling fT and g the long-run growth rate of traded goods production. Let S be the accumulated stock of foreign liabilities and suppose that RS is a reasonable estimate of their Iernandez-Arias and Montiel 65 future average service.' Under these assumptionis a solvency-based creditwor- thiness index can be constructed: (5) C = a(R -g)S/T where a is an arbitrary constant to base the index.' The index C, represents the ratio of the stock of external liabilities outstanding at date t to the projected present value of the resources available to service those liabilities from that date forward, ex- pressed relative to the same ratio during the base period. Thus C measures credit- worthiness in relative terms. An increase in this index has adverse implications for creditworthiness, and thus for the sustainability of external finance. A simple, alternative index could be based on the extent to which current capacity to pay meets short-term obligations, gauged by a liquidity-based ratio such as C/ = a'R'S/T, where R'is a short-term interest rate. Although this index lacks the theoretical foundations underlying C, it provides an interesting bench- mark. An even simpler alternative can be constructed by expanding the conven- tional debt-export ratio to include all external liabilities and all traded-goods productioll. In this case the index could he writteni as CD) = a'SIT. The three indexes are plotted in figure 2a for an aggregate of capital-inflow recipient countries. For predicting the level of the indexes in future years, projections of the pro- duction of traded goods and interest rates are nteeded. In figures 2a and 2b, the level of traded goods. T, is projected to grow at tlle rate g observed in the period 1989-93. The long- and short-term interest rates, R and R', respectively, are obtained from the implied forward rates of the maturity structure referred to above. We note four main points in comparing C to C' and Cl' (figure 2a). First, the relative evolution of the creditworthiness indexes is very sensitive to the evoll- tion of interest rates. The path of C tracks fairly closely that of market interest rates, both in the historical period, between 1983 and 1993, and in the pro jectionl period. Second, creditworthiness improved according to our preferred measure, even as capital flowed into developing countries, until end-1993, contrary to what the traditional index would suggest. In that sense this more refined index can better explain the surge in inflows. Third, creditworthlinless declined in 1994 and continues to do so in the projection period. Fourth, in spite of this decline, the index remains below its 1989 value throughout the projectioni period. This result reflects the fact that growth in T offsets projected increases in interest rates. We interpret this evidence as indicating that if the output of traded goods 5. This coincidence with the discount rate require% that rerurns on foreign investments adjust quickly to market conditions, as in the case of equity investments, floating-rate debt, or rolled-over short-term debt. 6. Note that we are assujmiig that the growth rate of 1is LilaffectUd b! changes in interest rates. This is a strong assumptionl, aJid to rhe extenit that it ftlils to hold thie Con1CILISIOns may be excesivel y optiniistic. Moreover, the index C. Is based ion fundamentals. It the fUndanrentals are themselves vulnera ble to perceptiions of noncreditworthi ness. creating scope for sell -utoilling rLH S, (itir optiiiiistic coclusions WoIild need to be qualified. 66 THF WORI-D BANK ECONOMIC REVIEW, VOL. 1(, NO. I Figure 2a. Creditworthiness Indexes with Constant Stocks of Liabilities Index 120- 100- 7Dl 80 Figure 2b. Creditworthiness Indexes with Growing Stock> of Liabilities Inclex 140- 10- 820 -40- 2 C -60 1989 1990 1991 1992 1993 1994' 1995' 1996* 1997* 1998' 1999* 2000' *Trojected rates. Vote. C= a(R-e)dtvo GLh= a'R'S/ C)=ea"S/ The constants a, a', and a" were determined such that C= Ct*= CnI= 100 at the start of 1990. Tis the level of traded goods projected to grow at the rate g observed in the period 1989-93. R and R' are, respectively, the long- and short-term interest rates. S is the stock of total foreign liabilities comprising debt and foreign equity (obtained by accumuilating equity net flows since 1970). In the projection period in figure 2a, S is held constant at its end-1993 level. In the projection period in figuire 2h, S grows at the rate observed in the period 1989-93. Data are for all developing countries as reported in World Bank (1994). Soturce. Authors calculations basecl on data from World Bank (1994). Fernindez-Arias and Montiel 67 grows at its estimated historic rate and market interest rates move as projected, the sustainability of the existing level of external liabilities will not be impaired by creditworthiness considerations, in that the creditworthiness index does not surpass values that were compatible with substantial capital inflows in the past. This interpretation suggests that creditworthiness considerations need not asso- ciate rising market interest rates with pressures for a reversal of capital flows and crisis. But can the inflow continue under such circumstances at rates comparable to those recently observed? To answer this question, an alternative measure of the index that incorporates growth in the stock of external liabilities S at the aver- age rate observed during the recent surge episode is used (figure 2b). These indexes assess whether creditworthiness would be impaired if inflow levels were to be sustained at levels on the order of those observed in recent years. Under these circumstances our preferred index C deteriorates over the projection pe- riod, but remains below its 1989 level by 2000. The implication is that consider- ations of country creditworthiness are unlikely to evolve in a way that will con- strain inflows in the near term. This does not imply, however, that portfolio considerations operating through the opportunity cost term Wn in equation 1 will not restrain such flows. Stocks and Flows Even if, as these results indicate, rising market interest rates do not necessar- ily portend a deterioration in C to critical levels, they do imply an increase in u. in equation 1, which itself has implications for the vector of flows F. These implications depend on how existing stocks S enter equation 1. We refer to a situation in which S enters equation 1 through the function C or W as one of stock adjustment, and refer to the alternative, in which all adjustment occurs through flows, as flow adjustment. To the extent that S enters C or W, even if the new inflows were purely a function of permanently improved domestic policies, it is unlikely that the mag- nitude of the initial flows would be sustained. The reason is that initial inflows would cause cumulative changes in stocks that would diminish the incentives for new inflows (by reducing C or increasing W, or both), and make the inflows a one-time event to some extent (see Fernandez-Arias 1995 for a formal analysis of the relative importance of flow-stock adjustment and the dynamics involved under expansion and contraction). For example, in the extreme case in which stocks are important for portfolio balancing reasons and domestic returns ad- justed for country risks are constant (F enters equation I only on the right-hand side), after the initial stock adjustment of foreign investors' portfolios is com- pleted, subsequent inflows would represent only the share of new saving de- voted to the acquisition of developing-country assets-that is, their magnitude would be limited by the rate of growth of foreign investors' overall portfolios. If stocks are important, the question of sustainability becomes one of how inflows can be expected to decrease under plausible scenarios, not whether in- 68 FIiF XokRI 1) RANK I (: )N(lI( RFVII K. Vol 10), N(O I flows will continue at their current levels. The answer depends on the perma- nence of the chaniges in the variables driving the inflows as well as on how much of the observed inflow in each country reflects an initial stock adjustment. Given the projected increase in international interest rates, capital inflows should fall for developing countries as a group, all other things being equal, continuing their estimated reduction during 1994. Nonetheless, in countries in which in- flows have primarily resulted from an improved domestic economic environ- ment that is expected to be maintained, there is no reason for the bulk of the stock adjustment to be reversed, even when external conditions change. Thus, although flows may taper off in such a case, reflecting both the completion of the initial stock adjustment and the change in external circumstances, a crisis is not likely to develop. If, instead, the contribution of domestic factors has been relatively minor, or even negative, and inflows have thus reflected primarily lower foreign interest rates, the stock adjustment canl be expected to be reversed if and wheil foreign assets become more attractive. So far, the only evidence on the empirical role of stock adjustment in the current inflow episode has been provided by Fernaindez-Arias (I 995), who found no evidence that flows responded to accumulated stocks. The importance of this issue for the prospective magiitude of postsuirge inflows and the likelihood of crisis warr'anits more research. Speed ol' Adustment The third and final component of the sustainability issue concerns the speed with which a desired stock reversal can be effected by external creditors. In equation i, adjustmenits are assumed to be costless and therefore instantaneous. But in practice the speed of adjustimient depends on the ease with which such creditors can liquidate their positions. In this regard the current inflow episode differs from thie previous o(.m. On the one hand, the bonds and equities acquired by external creditors in the current episode are more easily liquidated than syn- cilcated bank loaiis. Even FLi cain be liquidated effectively by borrowing domes- tically and transferring the funds abroad, particularly if outflows have been lib- eralized, as has been common in debtor couLitries during recent years. On the othier hand, the assets acquired by external creditors in the present example are denomlinated in domestic currency in many cases. This characteristic enhances liquidity while reiiderinig the foreign-currency value of such assets susceptible to capital taxation thrlougil their exposure to devaluationl. With assets that are rela- tively liquid and denomiiinated in domestic currency, portfolio adjustments are likely to be effected rapidly in response to niew inlforimation. V. PoucY RESPONSES The question of an appropriate policy response has received substantial at- tention, and the menu of policies considered has been extensive (see, for ex- ample, Calvo, Leiderman, and Reinhart 1993a; Schadler and others 1993). The Ferndndez-Arias and Montiel 69 desire to counteract the pressures for exchange rate appreciation in the face of substantial net capital inflows has typically led to very active Central Bank in- tervention and rapid increases in international reserves. Policies motivated by the desire to ameliorate this impact of capital inflows on the external compo- nent of high-powered money include: * Direct intervention to reduce gross inflows, by imposing controls or taxes on capital imports. * The removal of restrictions on capital outflows to reduce net inflows. * Trade liberalization, intended to switch spending from domestic to foreign goods and thus increase the trade deficit. * Increased exchange rate flexibility. In the last case the central bank fails to satisfy all of the demand for high- powered money created by capital inflows, allowing some of that demand to be reflected in an appreciation of the domestic currency. This could be accom- plished, for example, by allowing the currency to move within a band. An alternative approach is to accept some increase in the external component of the monetary base, but to counteract the potential effects of such an increase on domestic aggregate demand by using the conventional tools of macroeconomic policy, including tight fiscal policy and restrictive monetary policy, in the form of sterilized intervention or increases in marginal reserve requirements.' The first set of policies is aimed at reducing net inflows. If inflows have an exter- nal cause, these policies can be seen as general-purpose policies that attempt to reduce the size of the shock disturbing the economy. The other policies are likely to have feedback effects on the level of net inflows, however. Tight fiscal policy would reduce inflows by easing pressures on domestic interest rates and the trade deficit, while restrictive monetary policy would tend to increase inflows. The rest of this section examines how the nature of the appropriate policy response is affected both by the causes of the inflows and by the economic char- acteristics of the recipient country. Microeconomic Distortions Worsened by Exogenous Changes in Capital Inflows Consider first the case in which new capital inflows triggered by exogenous events aggravate the negative welfare consequences of a preexisting domestic distortion. A first-best policy response is to remove the distortion and absorb the capital inflow. Consider, for example, the case of improperly priced government deposit insurance. It may be impossible for the government to credibly eliminate such insurance. If this is the case, the insuranice should be priced properly to avoid subsidizing excessive risk taking (financed by both foreign and domestic deposits) on the part of depository institutions. Removal of the distortion would 7. Unsterilized intervention is nor included as a policy response hecause it represents tihe status quo, and thus reflects a passive policy stance. 70 THF WORI D A\NK F,(TON M1( RIXHIV. V0 10, NO I have been the prescribed policy even without the inflow. But if the first-best policy is infeasible, then direct interventionl in the form of capital controls or taxation to reduce the inflow emerges as a possible second-best policy response. Another important distortion emanates from the imperfect enforceability of cross-border contracts underlying country risk. An increase in foreign liabilities makes capital rationinig and debt crises more likely. The increase in the prob- ability of such events represents a cost that is external to domestic borrowers to the extent that other domestic agents share such costs, either through the ac- tions of external creditors or through the socialization of losses through the domestic political system. Such a borrower would thus have an incentive to attract too much foreign capital. This situation appears particularly relevant in countries that are close to their foreigin capital carrying capacity. In this case the distortion cannot be removed to any substa ntial extent, which again leads to a second-best approach to the problem. If an excessive level of foreign indebted- ness is directly caused by this distortion, a Pigouvian tax on capital inflows or equivalent capital control maya yield the required lower level of capital inflows and achieve the first-best outcome (since in this case the policy acts directly on the source of the distortion). Inflow s Indu ced by C/banges in Microcconoinic Distortions Excessive capital inflows can be induced by introducing new microeconomic distortionls or removinig old ones, for example, removing constraints on inflows. When inflows are triggered by the introduction of new distortions, the first-best policy response is to remove the distortions. (This point is made by Corbo and Hernandez 199.3.) The domestic distortionis most frequently mentioned in the role of attracting capital inflows are incredible trade liberalizations and price stabilizations. The solution to this type of distortion depends on the reason cred- ibility is absent. If it is absent because of the failure to set policy fundamentals (typically the fiscal deficit) on a sustainable parh, the solution is to adjust the fundamiienitals to attain credibility. If, however, such adjustments have been un- dertaken and credibility remainis elusive, theni direct intervention in the form of capital controls may again represent a second-best alternative. There is an obvi- ous analogy here to the use of wage and price controls in heterodox stabilization programs, in whichi adjustment in the fundamentals is complete, but lack of credibility or iniherenlt wage-price inertia threatenis to derail the stabilization prograin. Capital inflows can also arise because of the removal of distortions or coIn- straints. Microeconomic examples include the lifting of capital controls, the re- moval of barriers to direct foreign investment, and measures to enhance access to creditor-country finanicial markets. In addition, the adoption of a comprelhen- sive package of credible stabilization policies accompanied by liberalizing policy reforms can be thoug ht of as the comprehensive removal of widespread distor- tions. To the extent that such policies restore a country's creditworthiness, for example, they have the effect of removinig a prospective tax on its creditors. Fernindez-Arias and Montiel 71 In the absence of additional distortions, the removal of distortions constrain- ing capital inflows would move the economy to a nondistorted Pareto optimum, and thus improve welfare. In general, a capital inflow associated with the wel- fare-enhancing removal of distortions, whether in specific markets or as part of a generalized package of policy reforms, does not call for countervailing policies on microeconomic grounds. If other distortions are present, however, the out- come may be ambiguous, as second-best theory would predict. A preexisting distortion may be part of a second-best policy package, and removing it may result in a reduction in welfare when capital flows in. For example, as noted above, capital controls or taxes on external borrowing may be optimal in the presence of borrowing externalities arising from country risk considerations. Removing the policy "distortion" would induce capital inflows associated with overborrowing and thus produce an inferior welfare outcome. In such cases the correct policy stance is to retain controls. Capital Inflows and Macroeconomic Equilibrium We are left with the issue of macroeconomic instability-the question of how to use policy to preserve the macroeconomic equilibrium in the face of a foreign real interest rate shock. The first point to make with respect to macroeconomic policy goals is that it may prove optimal to leave policy unchanged. The shock will typically be expansionary. This may not be true if the recipient country operates a freely flexible exchange rate regime, as discussed below, but few of the countries that have been the recipients of the recent surge in capital inf]ows fit this description. Difficulty arises in the case of an economy operating at full capacity that seeks to preserve price stability. What are the policy options in this case? First, note that in the absence of any policy response the magnitude of the effect of a given fall in foreign real interest rates on domestic aggregate demand is likely to depend on whether the reduction is widely perceived to be temporary or permanent-that is, whether a fall in short-term rates is matched by a fall in long-term rates. The reason is that the capitalization of future income streams will depend primarily on whether long-term rates fall. A temporary reduction in foreign short-term rates may be associated with a capital inflow, but such an inflow is likely to be short-lived and perceived as such. Because it has little effect on domestic demand conditions, it creates no need for a stabilizing policy response. If the change is perceived as permanent, the full panoply of policy options described at the beginninig of this section is potentially relevant. The most direct option is to attempt to limit the size of net inflows arising from portfolio reallo- cations. To this end, controls on gross inflows could be introduced, in the form of ceilings or taxes, explicit or implicit, on foreign borrowing or on foreign direct investment. But it has been argued that this policy is not feasible because these limitations are always circumvented. Although it can be argued that even then the policy may be effective as long as tax avoidance is costly because it 712 THF WORiL[> hANK FCONONI( R[VIF\', Vol 1(, NO I reduces the return to investors, the social cost resulting from the attendant inef- ficient financial intermediation may disqualify this policy. More important, however, although capital controls could conceivably be a first-best solution if they respond to the microeconomic distortion directly in- ducing the capital inflows, or a second-best solution in circumstances such as those described above, capital controls are hard to justify in other cases. If the problem is macroeconomic in nature, the imposition of effective capital con- trols means introclucinig a niicroeconomic distortion. Macroeconomic stability may be preserved, but the costs of the distortion would remain. It would clearly be preferable to maintain stability without introducing a distortion by relying on more traditionial tools of stabilization policy. As in the case of microeconomic distortions described above, justification for capital controls would require a second-best argument based on the ineffectiveness of such tools (and relative effectiveness of controls) or on the high costs of employing them relative to the costs of the distortions introduced by controls. Alternatively, gross outflows could be promoted by liberalizing capital outflows. Assuming no other distortions, liberalization would be desirable even in the absenice of a foreign finanicial shock. Moreover, the argument that it is not feasible to impose controls applies to this case, too, and implies that outflows are already de facto liberalized. Even if effective, outflow liberal- ization could be counterproductive. Because limitations to capital repatria- tion are a concern to foreign investors, their removal is equivalent to the removal of a tax on foreign investment. Consequently, outflow liberaliza- tion will lead to increased gross inflows, which may more than offset the direct effect on increaised gross outflows. Current accouLnt liberalization, by contrast, many not cause the balance of payments to deteriorate, since under plausible circumstances liberalization may cause domestic saving to increase and (less plausibly) investment to decline (see Ostry 1991). Consequently, liberalization-of either the capital or the current accounts-may not relieve the upward pressure on the mon- etary base emanating from capital inflows. If the net inflow is not prevented from materializing through these means, a case can be made for undertaking a stabilizing macroeconomic policy response. However, the way in which the foreign financial shock is transmitted to domes- tic aggregate deriand-and thus the nature of the macroeconomic problem cre- ated by the shock-as well as the set of feasible macroeconomic policy responses is likely to differ from country to country. A key factor determining this response is the exchange rate regime. Under fixed exchange rates an autonomous capital inflow driven by a reduction in foreign interest rates leads to inflation and lower real domestic interest rates if monetary policy is passive and limited to unsterilized intervention. To avoid this outcome, the authorities could switch to sterilized interventioni. This policy has the appeal of supplying foreigners with the domestic interest-bearing assets that they demand while still adhering to a domestic money supply target for stabili- Ferndndez-Arias and Montiel 73 zation purposes. (Reisen 1993 has been a forceful advocate of this policy.) Con- trary to what is sometimes asserted, sterilization does not necessarily imply that the inflow will be perpetuated, since the inflow will end once portfolio composi- tion has adjusted to accommodate rate-of-return differentials.8 Sterilization, however, is not a panacea. It may not imply the infinite perpetu- ation of the inflows, but it will tend to magnify the size of the cumulative inflow. Moreover, it may not insulate the domestic economy. If domestic financial assets are regarded as imperfect substitutes by foreign investors and if the instrument used to sterilize is not demanded by foreign investors, then domestic portfolio equilibrium will require an adjustment in relative rates of return among domestic assets. Even if it insulates successfully, sterilization cannot be a permanent solu- tion-as long as the inflow persists, the central bank will be exchanging high- yielding domestic assets for low-yielding foreign ones, and this policy may have important fiscal implications. Financing the quasi-fiscal deficit that arises from such asset swaps would require a permanent transfer from the government to the central bank that is passed on to foreigners in the form of returns that are ele- vated relative to what they could earn at home. Even if fiscally feasible, such a policy is unlikely to prove palatable for very long. Finally, sterilization may turn out to be infeasible even in the short run if capital mobility is sufficiently high. Alternatively, a tighter monetary policy could be pursued by increasing mini- mum reserve requirements on banks' short-term foreign liabilities. These amount to a tax on foreign borrowing, which, like other taxes on capital inflows, may be difficult to implement. A specific problem with this approach is that it is likely to redirect capital inflows to domestic borrowers through channels other than the domestic banking system-such as through markets for equity and real es- tate. If this disintermediation is effective, the macroeconomic stabilization prob- lem would remain. The scope for circumventing the domestic banking system depends on the menu of domestic assets available to foreigners and thus on the degree of sophistication of the domestic financial system. Under flexible exchange rates the foreign interest rate shock will result in an appreciation of the domestic currency and possibly a small decrease in domestic interest rates, which would result, with a fixed money supply, from the price- level effects of a nominal appreciation. The external interest rate shock may prove to be contractionary, as expenditure switching adversely affects the de- mand for home goods. Stability in this case would require a monetary expan- sion, resulting in a combination of domestic interest rates that are lower than they would have been without the shock, but higher than under fixed exchange rates and a passive monetary policy. In addition, the exchange rate would ap- preciate relative to what it would have been without the shock, but depreciate relative to what it would have been without monetary expansion. This outcome is the basis for the policy advice proffered by both by Calvo, Leiderman, and Reinhart (1993b) and Schadler and others (1993), advocating a 8. This result can be derived from simple portfolio models. 74 ii: I ()WIR) BANK I( B ONO I(I RFV FI';. \'()i I(. N(. I role for exchange rate appreciation in adjusting to the external interest rate shock. Again, however, this advice may not be universally applicable. Countries that rely on the exchange rate as a nominal anchor will be reluctant to move the rate for fear of eroding the credibility of the peg. In addition, the degree of real appreciation may exceed that which would occur with a fixed peg, and thus this policy may hurt competitiveness. If these constraints are binding, the monetary policy options available are those outlined previously. These considerations suggest that policy may need to be prepared to accom- modate a reduction in domestic interest rates with an unchanged nominal peg. If so, the set of remaining policy options is narrow indeed. To preserve macroeconomic stability under such circumstances, the induced increase in pri- vate absorption would have to be offset through tighter fiscal policy. \VI. SUtMMARY AND CONCLUSIONS The current capital inflow episode represents a sharp break from the experi- ence of the debt crisis of the 1980s. The magnitude of flows nearly matches that which preceded the debt crisis. Although this surge constitutes a welcome relief from the constraints of credit rationing for many countries, it also poses struc- tural and macroeconomic policy challenges. The structural challenge is to en- sure that the resource inflow is efficiently used in order to avoid a repetition of the debt crisis. Although certain characteristics of the current inflows are reas- suring in this regard, potential disruption from several distortions implies that a laissez-faire stance is not necessarily warranted. Moreover, though capital in- flows may represent the outcome of a favorable external shock from the per- spective of indebted developing countries, their effect on macroeconomic stabil- ity may call for a policy response on these grounds as well. Why have capital flows to developing counitries resumed on a large scale? In the aggregate the role of foreign interest rates as a push factor driving capital inflows and determining their magnitude is well established by the systematic empirical work undertaken on this issue. At the same time, theoretical consider- ations suggest that the creditworthiness of the recipient country must have played an important role in determining both the timing and geographic destination of the new capital flows. We know little about the relative weights to assign to domestic and foreign factors in attracting capital to individLial countries and consequenitly even less about the role of specific types of domestic shocks. The existing evidence also sheds little light on the roles of domestic or external struc- tural factors. Our analysis suggests that this type of information is crucial for designing policies. Specifically, more country-specific information is required about the possible role of domestic microeconomic distortions in motivating these inflows and channieling them to the final borrowing sector. This discussionl makes clear that sustainability has an important endogenous component. The loss of creditwortliness due to a deterioration of the domestic policy stance is suLfficienlt to stop inflows quickly, and given the nature of stock Fernandez-Arzas and Montiel 75 adjustment, the liquidity of the assets acquired by external creditors, and their vulnerability to exchange rate changes, inflows are likely to be replaced by sub- stantial outflows or an outright balance of payments crisis. Recent events in Mexico provide strong support for this assertion. Even if creditworthiness is retained, however, the early level of inflows is unlikely to be sustained. The nature of shock adjustment would make the level of inflow diminish over time, even with stable external financial conditions, and, more so, the favorable for- eign financial shock that triggered the episode may not persist. Whether the outcome is a gradual reduction in flows since the early 1990s or an actual rever- sal depends on the path followed by foreign interest rates as well as on the role of stock adjustment. The key gap in knowledge concerns how large the tempo- rary stock adjustment component of the recent inflows has been relative to the permanent flow component. What are the implications for policy in the recipient countries? Establishing the feasibility of controls that would prevent the arrival of capital inflows is problematic and likely to prove country-specific. A case for direct intervention as a first-best policy can be made only when the negative welfare consequences of a distortion that cannot be removed arise from induced external borrowing. This circumstance is likely to apply in the context of country-risk externalities and may also apply in the presence of "incredible" reforms. In both situations, however, the appropriate intervention is a Pigouvian tax (or equivalent control) rather than a ban on capital inflows. Beyond this case, direct intervention would have to be based on second-best considerations, either on microeconomic or macroeconomic grounds. On the other hand, the receipt of capital inflows may strengthen the case for the removal of certain microeconomic distortions, either because they aggravate the costs of such distortions or because they ease the perceived constraints (tvpically balance of payments constraints) that originally motivated their adoption. To the extent that capital inflows are permitted to materialize, the desirabil- ity of foreign exchange intervention depends on the requirements for macroeconomic stability. Either competitiveness considerations or use of the exchange rate as a nominal anchor in the context of a stabilization program may preclude nominal appreciation. If not, then permitting a (temporary) ap- preciation of the nominal exchange rate bv restricting the scale of foreign ex- change intervention-perhaps in the context of an exchange rate band-will dampen, and may reverse, the expansionary effect of the foreign interest rate shock on domestic aggregate demand by appreciating the real exchange rate and possibly raising the domestic interest rate. This outcome will be desirable if domestic macroeconomic conditions are such that policymakers seek to avoid stimulating aggregate demand. Alternatively the authorities can avoid aggre- gate demand stimulus with a fixed exchange rate through sterilized foreign ex- change intervention. But this policy is feasible only if capital mobility is imper- fect. The higher the degree of capital mobility, the larger will be the accumulation of reserves associated with a policy of sterilizationi. This policy has associated 76 THE WO ORLD BANK Y( (NOM I RF\VII W. V( L. ]), N(. I quasi-fiscal costs, since the central bank exchanges high-yielding domestic as- sets for low-yielding reserves, and the magnitude of these costs will be greater the higher the degree of capital niobility and the larger the gap between domes- tic and foreign rates of return. Moreover, even if successftil, this policy may not insulate the economy from the expansionary effect of the foreign shock if substi- tution among domestic assets is imperfect and the asset demanded by external creditors is not that used in intervention. If sterilization is incomplete, the implication of the inflow is an expansion in the monetary base. Monetary expansion can still be avoided by a commensurate reduc- tion in the money multiplier achieved through an increase itn reserve requirements. In this case quasi-fiscal costs are avoided through implicit taxation of the banking system. The economic implications of this tax will depend on how the tax burden is ultimately shared among the banks, their depositors, and their loan customers. Whether such measures can avoid an increase in aggregate demand depends on the structure of the domestic financial system, which determines the scope for disintermediation. Finally, if domestic monetary expansion is not avoided, or if an expansionary fin.ancial stimulus is transmitted outside the banking system, the sta- bilization of aggregate deemand will require a fiscal contraction. The key message is that choices confront niacroeconomic policymakers at each step in this progression. Not only the intenided effect on aggregate demand, but also the feasibility and relative desirability of alternative macroeconomic policy packages to achieve that effect will be functions of country circumstances. Relevant considerationis include the economy's level of capacity utilization, the identity of its nominal anchor, the sterilization tools available to the central bank, the degree of capital mobility, the finaticial healthi of domestic banks, the sophistication of the financial system, and the flexibility of fiscal policy, among others. In view of the multiplicity of factors that should in principle influence the response of macroeconomic policies, no single combination of policies is likely to be optinial in all cases. REFEREN(T S The word "processed" describes inforimially reproduced works that mav not be com- monly available through library systems. Bacha, Edmar L. 1993. "Selected Interinational lPolicy Issues on Private Market Financ- ing for Developing Countries." Paper for the lUnited Nations Conference on Trade and D)evelopment (UNC(IAD). United Nations, New York. Processed. Calvo, Guillermo. 1989. "Inicredible Reforms." In G. A. Calvo, R. Findlay, P. Kouri, and J. Braga, eds., Debt, Stabilization, anzd Development. Cambridge, Mass.: Basil Blackwell . Calvo, Guillermo, and Carlos A. Vegh. 1991. "Exchanlge Rate Based Stabilization un- der Imperfect Credibility." INI Working Paper 91-77. Iliternational Monetarv Fund, Research Department, Washiingtoil, D.C. Processed. Ferndndcez-Arias and Montiel 77 Calvo, Guillermo, Leonardo Leiderman, and C.armen Reinhart. 1993a. "The Capital Inflows Problem: Concepts and Issues." International Monetary Fund, Research Department, Washinigtoin, D.C. Processed. .1993b. "Capital Inflows and Real Exchange Rate Appreciation in Latin America: The Role of External Factors." International Monetary Fund Staff Papers 40(March):108-51. Corbo, Vittorio, and Leonardo Hernandez. 1993. "Macroeconomic Adjustment to Capital Inflows: Rationale and Some Recent Experiences." World Bank, Interna- tional Economics Department, Washington, D.C. Processed. Dooley, Michael, Eduardo Fernindez-Arias, and Kenneth Kletzer. 1996. "Recent Pri- vate Capital Inflows to Developing Countries: Is the Debt Crisis History?" The World Bank Econiomic Review 10(1):27-49. Fernandez-Arias, Eduardo. 1995. "The New Wave of Private Capital Inflows: Push or Pull?" Journal of Development Econzomics (December). Fernandez-Arias, Eduardo, and Peter J. Montiel. 1995. "The Surge in Capital Inflows to Developing Countries: Prospects and Policy Response." wls 1473. World Bank, International Economics Department, Washington, D.C. Processed. IMF. Various years. International Financial Statistics. Washington, D.C.: International Monetary Fund. Montiel, Peter J. 1993. "Fiscal Aspects of Developing-Country Debt Problems and [)DSR Operations: A Conceptual Framework." wi's 1073. World Bank, International Eco- nomics Department, Washington, D.C. Processed. 1995. "The New Wave of Capital Inflows: Country Policy Chronologies." Oberlin College, Department of Economics, Oberlin, Ohio. Ostry, Jonathan. 1991. "Trade Liberalization in Developing Countries, Initial Trade Distortions, and Imported Intermediate Inputs." Internationial Monetary Fund Staff Papers 38(September):447-80. Reisen, Helmut. 1993. "The Case for Sterilized Intervention in Latin America." Organisation for Economic Co-operation and Development (OECD), Paris. Processed. Schadler, Susan, Maria Carcovic, Adam Bennett, and Robert Khan. 1993. Recent Ex- perience with Surges in Capital Inflows. IMIF Occasional Paper 108. Washington, D.C.: International Monetary Fund. World Bank. 1994. World Debt Tables 1994-9 5: External Finance for Developing Coun- tries. 2 vols. Washington, D.C. A SYMPOSIUM ON FERTILy IN SUB-SAHARAN AFRICA Introduction: Fertility in Sub-Saharan Africa Martha Ainsworth During the 1980s the population of Sub-Saharan Africa grew at a rate of 3.1 percent per year, the highest of any developing region (World Bank 1993). The population of South Asia, the developing region with the next highest rate, grew at 2.2 percent annually. Recent demographic surveys in Sub-Saharan Africa found that the average total fertility rate (TFR)-the number of children a woman would have in her lifetime at prevailing age-specific fertility rates-is generally be- tween six and seven children per woman. Child mortality has declined steadily since World War II, but infant and child mortality remain relatively high (Hill 1990). In fifteen Sub-Saharan African countries, the infant mortality rate ex- ceeds 100 per 1,000 live births, and in four countries the rate is greater than 140 per 1,000 (World Bank 1995). There are signs of fertility decline in a few coun- tries (Botswana, Kenya, and Zimbabwe), but even in these cases total fertility is relatively high at five or more children per woman. Economic growth in Sub-Saharan Africa has lagged behind population growth. Between 1965 and 1988, the gross national product (GNP) per capita grew by only 0.2 percent annually for the region; during the 1980s, average income per capita declined (World Bank 1990, 1993). Levels of human capital in the form of schooling and other training are low, and school enrollment rates have actu- ally fallen in many countries. Although conditions vary a great deal, many coun- tries, if not most, face the prospect of a rapidly growing labor force with low levels of human capital. Changing this scenario will require not only policies that help restore economic growth but that also enable families to have fewer children and to invest more in the quality of each child. To this end, most Afri- .can countries have adopted or have endorsed public provision of subsidized family planning services, which provide families with not only the means to implement their fertility preferences but also to improve maternal and child health. How- ever, levels of modern contraceptive use in all but a handful of Sub-Saharan African countries are still below 10 percent. The persistent high levels of fertility and low levels of contraceptive use in most Sub-Saharan African countries have fostered a lively debate on two ques- tions relevant to the design of population and human resource policies. First, is Africa different from other developing regions in terms of the factors influenc- ing the demand for children? If it is sufficiently different, will the policies or factors that have affected the decline of fertility in other developing regions be Martha Ainsworth is with the Policy Research Department at the World Bank. O 1996 The Interiational Bank for Reconstruction and Development /1HE WORLD BANK 81 82 THE WORLI) BANK FCONOMIC REVIEW. VOL. 10. NO. I effective in Sub-Saharan Africa? Second, are these outcomes due primarily to low levels of economic development that encourage large families or to insuf- ficient provision of family planning information and methods? The articles in this symposium speak to both of these issues. 1. Is AFRICA DIFFERENT? Both the demographic and anthropological literatures point to important cul- tural institutions in Sub-Saharan Africa that encourage high fertility and spread the costs and benefits of children beyond the couple in question. Caldwell, Orubuloye, and Caldwell (1992), for example, mention the importance in tradi- tional society of maintaining a lineage and leaving descendants, the separate budgets and decisionmaking between spouses, the spreading of child costs ac- ross the extended family through child fostering, and communal land tenure systems that favor large families. These institutions arose in an environment that favored high fertility. Are they impediments to demographic change? The first two articles in the symposium (Ainsworth, Beegle, and Nyamete and Benefo and Schultz) show that despite these cultural influences, African fertility and contraceptive use are sensitive to policies associated with fertility decline elsewhere in the world. Perhaps one of the strongest associations world- wide is the negative relation between female schooling and fertility. Ainsworth, Beegle, and Nyamete confirm this relationship in a cross-national study of in- dividual fertility determinants in fourteen Sub-Saharan African countries. Their study is important because it controls for other exogenous correlates of fertility and contraceptive use to isolate the relation with female education at the microlevel, and it does this in a comparable way across a large group of coun- tries. Further, the authors distinguish between the effect of the early and late years of primary schooling, a point that has received a great deal of attention in bivariate analyses of female schooling and fertility. The most intriguing results are the important differences found across countries in the magnitude of the negative schooling-fertility relationship and in the relative impact of female and male schooling. What might explain these differences? The authors point to several possibilities-underlying differences in the quality of schooling, the la- bor market, child health, family planning programs, and the status of women. The data sets used did not permit an analysis of these factors, but the results suggest avenues for additional research, which could yield important policy in- sights. In the meantime the results of this article and others speak to an urgent need to raise the stunningly low levels of completed schooling among African women in order to lower fertility and improve child quality (see Montgomery, Kouame, and Oliver 1995). High levels of child mortality are also thought to be an impediment to fertil- ity decline in Sub-Saharan Africa. Benefo and Schultz show that high levels of child mortality in Cote d'lvoire and Ghana are resulting in higher fertility through a child "replacement effect." The estimated effect is relatively small, but it might Ainsworth 83 be expected in an environment where child mortality rates are very high and have not fallen below the critical threshold at which a response of lower fertility might be anticipated. C6te d'lvoire and Ghana are adjacent countries with many cultural and geographic similarities, yet the determinants of fertility and child mortality are remarkably different in the two countries. This belies their differ- ent colonial experience and economic and social policies since independence. Ghana, for example, has equalized educational opportunities between men and women to a far greater extent than C6te d'lvoire. 11. FAMILY PLANNING PROGRAMS ANI) THE DEMAND FOR CHILDREN A recent study of the prospects for fertility decline in Sub-Saharan Africa concluded that there are small groups of women who want fewer children and who do not have easy access to family planning (van de Walle and Foster 1990). But, by and large, desired family size is still high-between six and nine children per woman. Thus, lowering fertility and raising contraceptive use will depend both on lowering the demand for children and increasing the availability of fam- ily planning services and information. The last two articles (Feyisetan and Ainsworth and Thomas and Maluccio) assess the relative importance of family planning services and the factors af- fecting the demand for children in determining contraceptive use. The case stud- ies analyze two countries at different stages in the demographic transition. Ni- geria is the most populous country in Africa with limited female schooling and a nascent family planning program. Fewer than 5 percent of women are using modern contraception. At the other end of the spectrum is Zimbabwe, where extensive public investments in female schooling, family planning, and health infrastructure since independence in 1980 have already brought about a modest fertility decline. More than a third of Zimbabwean women are using modern contraceptive methods, one of the highest rates in Sub-Saharan Africa. By linking women to the results from parallel service availability and "situa- tion analysis" surveys, the authors attempt to control for the woman's charac- teristics as well as the quality, availability, and (in Nigeria) the price of family planning methods in their analyses of contraceptive use. This approach has re- cently been used to examine the impact of public health services on morbiditv and child nutrition (Alderman and Lavy 1996) and the choice of medical pro- vider (Mwabu, Ainsworth, and Nyamnete 1993) in a number of African coun- tries. Female schooling is a pervasive and very strong correlate of contraceptive use in both Nigeria and Zimbabwe despite their different stages in the demogra- phic transition. Feyisetan and Ainsworth find that the low availability of health and family planning services is constraining contraceptive use in Nigeria but that outpa- tient consultation fees are not. Broadly similar evidence emerges from Zimba- bwe where, in addition, Thomas and Maluccio highlight differences in the im- pact of family planning programs on behaviors of poorer and better-off women. 84 FHF WORLD BANK FCC0NOM\l RFVIEW, VOL. 10, NC). I They report that the system of community-based distributors is associated with higher adoption of contraceptives among better-educated women, whereas im- proved quality of distributors benefits the least educated more. The Nigerian and Zimbabwean studies are important reminders that different aspects of fam- ily planning services may be relatively more influential in raising contraceptive use at different phases of the demographic transition in Sub-Saharan Africa. They also make a strong case for strengthening the capacity to evaluate pro- gram interventions in Africa through random assignment and phased imple- mentation of services to experimental and control communities. Taken together, these four papers suggest that Sub-Saharan African fertility, although subject to unique cultural influences, can be expected to respond to many of the policies that have been found to lower fertility and raise invest- ments in children in other parts of the world. They also point to fruitful avenues of research on the extent to which observed differences across countries can be attributed to differences in public policies. REFERENCES Alderman, Harold, and Victor Lavv. 1996. "Household Responses to Public Health Services: Cost and Quality Tradeoffs." The World Bank Research Observer 1 l(1):3-22. Catdwell, John, 1. 0. Orubuloye, and Pat Caldwell. 1992. "Fertility Decline in Africa: A New Type of Transition?" Population and Developmnent Review 18(2):211-42. Hill, Althea. 1990. "Population Conditions in Mainland Sub-Saharan Africa." In George T. F. Acsadi, Gwendolyn Johnson-Acsadi, and Rodolfo A. Bulatao, eds., Population Growth and Reproductioni in Sub-Saharan Africa: Technical Analysis of Fertility and Its Consequences. A World Bank Symposium. Washington, D.C. Montgomery, Mark, Aka Kouame, and Raylynn Oliver. 1995. The Tradeoff between Number of Children and Child Schooling: Evidence from C6te d'lvoire and Ghana. LSMS Working Paper 112. Washington, D.C.: World Bank. Mwabu, Gerinano, Martha Ainsworth, and Andrew Nyamete. 1993. "Quality of Medi- cal Care and Choice of Medical Treatment in Kenya." journal of Human Resources 28(4):838-62. van de Walle, Etienne, and Andrew Foster. 1990. Fertility Decline in Africa: Assess- ment and Prospects. World Bank Technical Paper 125. Washington, D.C. World Bank. 1990. World Development Report 1990: Poverty. New York: Oxford University Press. . 1993. World Development Report 1993: Investing in Health. New York: Ox- ford University Press. . 1995. World Development Report 1995: Workers in an Integrating World. New York: Oxford University Press. The Impact of Women's Schooling on Fertility and Contraceptive Use: A Study of Fourteen Sub-Saharan African Countries Martha Ainsworth, Kathleen Beegle, and Andrew Nyamete This article examines the relationship betuween female schooling anid two behav- iors-cumulative fertility and contraceptive use-in fouirteen Sub-Saharan African countries where Demographic and Health Surveys (OHS) have b7een coynducted since the mid-19SOs. Average levels of schooling amiong women of reproductitie age are very low, from less than tuvo years to six. Controlling for background variables, the last years of femnale primary schooling have a negative relation with f ertility in aibout half the countries, while secondary schooling is associated with substantially lower fertility in all couintries. Female schooling has a positive relationship with conitracep- tive use at all levels. Among ever-married wvomen, husband's schoolinig exerts a smaller effect than does fem"ale schooling on contraceptive use and, in almost all cases, on fertility. Although the results suggest commonalities anmong these Sub-Saha ran coun- tries, they also reveal intriguing international differcnc-es in the imzpact of female schooling, which might reflect differences in the quality of schooling, labor mnarkets, and family planning programs, amiong others. There is considerable debate in the literature as to whether high fertility and high desired family size in Africa are caused by low levels of economic devel- opment that favor large families (see World Bank 1984, 1986), or by unique cultural features (see Caldwell and Caldwell 1987, 1990). Without denying the possibility that cultural traits may contribute to higher demand for children in Africa than in other developing regions, most studies have found differentials in current or total fertility by socioeconomic class, even in high-fertility countries Martha Ainsworth is with the Policy Research Department at the World Baink; Kathleen Beegle is with the Department of Economics ar Michigan State University; and Andrew Nyamete is with AIDS Control and Prevention (AIDSt AP). This article was writtenl as background for the research proiect on "The Economic and Policy Determinants of Fertility in Sub-Saharan Africa," financed hy the World Bank Research Committee (RPo 67691) anid sponsored by the Africa Technical Department and the Policv Research Department of the World Bank. The authors are grateful to the World Bank Research Committee and the Africa Region for financial assistance and to Johii (aidwell, Cynthia Cook, David Cornelius, joy de Bever, Barbara Herz, Cynthia Lloyd, Gora M'Boup, Amolo Ngwerii, Lant Pritchett, Fred Sai, Guillerme Seladcek, David Shapiro, Paul Shaaw, Duncan Thomas, and Barbara Torrey for comments on earlier drafts. The authors also acknowledge Evina Akam, Clara Favorsev, Mouliamadou Gucye, Alice Kouadio, Katharine Namuddu, and others whii offered conrmente at disscmillatioil seminars in Camerooii, C'te d'lvoire, Ghana, Kenya, and Mali. A special debt of gratitude is cxtended to Susm1it;a Ghosh aiid KathY Burke for assistance. (© 1996 The International Bank for Reconstruction and Development / rHI: \XORI D BANK 85 86 THIL WORLD BANK FCONOMIC REVIEW. VOL. 10, NO. I (see Cochrane and Farid 1990 and United Nations 1987, for example). Among the factors thought to be the most conducive to the high demand for children are high child mortality and low levels of female schooling. This article estimates the relationships between female schooling and fertility and between female schooling and contraceptive use in fourteen Sub-Saharan African countries where Demographic and Health Surveys (DHS) were conducted from the mid-1980s to the early 1990s. Multivariate analysis of the determi- nants of fertility and contraceptive use allows a precise exploration of the rela- tionship with women's schooling, while controlling for variables like age, area of residence, wealth, ethnicity, and religious affiliation. Results across the coun- tries, for both national and subnational samples, are compared and form the basis for future research to explain international differences. Section I describes the posited relationships between women's schooling and fertility and between women's schooling and contraceptive use and summarizes the evidence to date on these relationships in Sub-Saharan Africa. Section II presents the empirical model and describes the data sets. Section III provides descriptive statistics-both from the data sets and from outside sources-on economic and demographic indicators for the fourteen countries under study. Section IV describes the results for determinants of fertility, and section V describes the results for the determinants of contraceptive use. Section VI re- views the results and proposes areas for additional research. 1. WO.MEN'S SC.HOOLING, FERTILITY, AND CONTRACEPTIVE USE Women's schooling is posited to result in lower fertility and, by inference, higher contraceptive use, through four main channels. W Wage effects. By raising the opportunity cost of women's time in rearing children, schooling raises the "price" of children (who are time-intensive) as well as the wage that women can earn in the work force. The wage benefits of schooling may also induce women to get more schooling, thereby delaying the onset of childbearing. This is likely to result in lower fertility and higher rates of female participation in the labor force. * Higher demand for child schooling. Women with more schooling may develop higher aspirations for their own children's schooling. These aspirations may lead them to have fewer children and to invest in more schooling per child. This is the quantity-quality tradeoff observed in other parts of the world but examined only recently in Sub-Saharan Africa (Kelley and Nobbe 1990; Montgomery, Kouame, and Oliver 1995). There are many initiating factors for this tradeoff, including the levels of wages and employment expected by graduates and the quality and price of schooling. * Lower child mortality. Women with more schooling are likely to be more effective in producing healthy children, which lowers child mortality. As the "wedge" between live births and surviving children is narrowed, couples find that they can have fewer children to reach a target number of surviving Ainsworth, Beegle, and Nvamete 8 7 children.' In a cross-national study, Schultz (1994) found that fully half of the effect of female schooling in lowering fertility was operating through its effect in lowering child mortality. More effective use of contraception. Educated women can learn about and use contraception more effectively than uneducated women, reducing the number of unanticipated pregnancies. It has been suggested that female schooling can indirectly raise fertility by improving maternal health, reducing pathological sterility, and reducing the du- ration of breastfeeding and its contraceptive benefits (Alam and Casterline 1984; Bongaarts, Frank, and Lesthaeghe 1984; C.asterline and others 1984; Cleland and Rodriguez 1988; Cochrane 1979, 1988; Jejeeboy 1992; and World Bank 1984). Female education is also thought to facilitate fertility decline by increas- ing the bargaining power of women, allowing them greater control over their destiny, and improving husband-wife communication (Jejeeboy 1992; United Nations 1987). There have been several studies of the differentials in aggregate measures of fertility between urban and rural areas and among women according to their level of schooling, using data from the World Fertility Survey (WFS) (Alam and Casterline 1984; Cleland and Rodriguez 1988; Cochrane 1988; Cochrane and Farid 1990; and United Nations 1987). The total fertility rate (TFR) is the num- ber of children a woman would have in her lifetime if she bore children accord- ing to current age-specific fertility rates. Table I presents differentials in the TFR by area of residence and female education for African countries that partici- pated in the more recent DHS. The TFR measured in these fourteen countries is very high-generally in the range of six to seven children. However, in every country it is significantly lower in urban than in rural areas. One reason for these urban-rural differentials is the concentration of women with secondary and higher levels of schooling in urban areas. Women who have completed primary schooling or who have some secondarv schooling univer- sally have a lower TFR than women without schooling. The differential between the fertility of women with primary schooling and those with no schooling is smaller and sometimes follows an unanticipated direction: in Burundi, Cameroon, Kenya, and Nigeria, women with some primary schooling actually have a higher TFR than those with no schooling. Note, however, that the TFR for those who completed the primary cycle is substantially lower than for women with no schooling in all cases, including Kenya and Nigeria. Results that indicate a posi- tive relation between some primary schooling and the TFR cast doubt on the effectiveness of less-than-complete female primary schooling in lowering fertil- ity, suggesting that schooling does not have a depressing effect on fertility until the secondary level (Cleland and Rodriguez 1988; Cochrane 1979, 1988; United 1. See Benefo and Schultz ( 1994) and Pitt (1995) for evidence of the relation between female schooling and child mortality in Sub-Saharafi Africa. 88 THE WORLD BANK tECONOMIC REVIEW, VOL. 1). NO. I Table 1. Total Fertility Rates for Women Age 15-49, by Residence and Educationz, Fourteen Sub-Saharan African Countries Education Residence Completed More than Countrn Year All Urban Rural None Primarv primary primary Tanzania 1991-92 6.3 4.0' 6.6' 6.5 6.4d 6.0 4.2 Uganda 1988-89 7.3 5.7 7.5 7.7 7.2d 7.3 6.7/5.lI Burundi 1987 6.8 5.3 6.9 6.8 7. - 5.5 Mali 1987 6.7 6.1 7.0 6.8 6.2 - n.a. Niger 199) 7.4 6.7 7.5 7.5 6.3' - Nigeria 1990 6.0 5.0 6.3 6.5 7.2d 5.6 5.1/4.29 Kenya 199.3 5.4 3.4 5.8 6.0 6.2w 5.0 4.0 Ghana 1993 5.5 4.0 6.4 6.7 6.1 - 4.7/2.9' Togo 1988 6.6 4.7 7.0 6.8 5.7 - 4.5 Zambia 199) 6.5 5.8 7.1 7.1 6.8 - 4.9 Zimbabwe 1988 5.5 4.1 6.2 7.0 6.0 - 3.8 Senegal 1992-93 6.0 5.1 6.7 6.5 5.7 - 3.8 Cameroon 1991 5.8 5.2 6.3 6.2 6.4 - 4.5 Botswana 1988 5.0 4.1 5.4 6.0 5.2d 4.6 3.3 n.a. Not applicable. - Not reported. Note: CoLuntries are listed in order of 1991 gross national product (GNP) per capita, froni lowesr to highest (see table 2). a. In Botswana, Kenya, Tanzania, Uganda, Zambia, and Zimbabwe, primary schooling includes seven years of instruction. For all other countries this figure is six years. Unless otherwise noted, primary refers to any primary schooling, including completed primary. b. The figure for Dar es Salaam is 4.0; for other urban maitiland, 5.6; for Zanzibar, 6.4. c. Mlainland onIy. d. Sonie primary. e. First figure is for middle school, second is for secondary and higher in Ghana and secondary 4 and higher in Ulganda. f. Any schooling (primary or more). g. First figure is for some secondary, second is for completed secondary and higher. Source: DHS COuLntry reports, available from IRD/Macro International, Inc., Calverton, Md. Following are countries and publicition dates of each report: Botswana (1989); Burundi (19881; Camerooi (1992); Ghana (1994);Keriya (1994); Mali (I1989); Niger(1993); Nigeria (1992);Senegal (1994):Tanzania (1993); Togo (1 989); Uganda (1989); Zambia (1993.1; arid Zimbahwe '1989). Nations 1987; World Bank 1984). However, these patterns are not apparent in the relation between women's education and contraceptive use: in African coun- tries covered by the DHS, women with more schooling are also increasingly likely to be practicing contraception (Castro Martin 1995). In fact, even bivariate studies of the correlates of contraceptive use conducted in the 1960s found that use increased with levels of male and female education (see Cochrane 1979, table 5.6). The comparison of aggregate measures of fertility and contraceptive use by socioeconomic status is a useful starting point for analysis, but has many short- comings. First, the two-way comparisons do not simultaneously control for other factors that influence fertility and contraceptive use. Income, for example, may be highly correlated with schooling but may have an opposite (positive) effect on fertility (Ainsworth 1989, 1990; Farooq 1985; National Research Council Aznstworth. Beegle, and Nvaniete 89 1993). Failure to control for these other correlated variables ma,v confound the independent effects of schooling, other policy variables, and exogenous factors like ethnicity. Second, the total fertility rate does not represent the completed fertility of any individual or cohort of women; rather, it is the number of children a woman would have over her reproductive lifetime if she were to bear children at prevailing age-specific fertility rates. Thus, cohort effects of schooling on the timing of births over the life cycle are not reflected in the comparisons of TFR for schooled and unschooled women. As a larger share of each successive cohort is educated, the relation between schooling and fertility for any given colhort is likely to change. Third, the policy objective is to lower aggregate measures of fertility by influ- encing individual decisions on the number of children. At the individual level, the policy objective is therefore to influence cumulative and completed fertility. Finally, if there are any nonlinearities in the relationship between policy vari- ables and fertility, they will not necessarily be reflected in, and cannot be studied with, aggregates like the total fertility rate (Anker 1985; Schultz 1992). For these reasons, multivariate analysis of individual data may yield greater insights on the likely impact of policies, such as female schooling, on cumulative fertility. A number of multivariate studies have examined the relation between female schooling and cumulative fertility in Sub-Saharan countries, controlling for the woman's age, area of residence, and sometimes her religion, ethnic group, and household income (see Ainsworth, Beegle, and Nyamete 1995, table 2). Six studies that controlled for household permanent income-using proxies such as husband's schooling, husband's income, household consuLIMptio1, land ownership, and cattle-found either a significant negative relation (Ainsworth 1989; Okojie 1990, 1991; Snyder 1974) or no relation (Anker and Knowles 1982; Farooq 1985) between women's schooling and cumulative fertility at low levels of female school- ing. At higher levels, female schooling was correlated with substantially lower fertility. Burafuta and Shapiro (1992) found a positive relation between primary schooling and fertility in Burundi, but secondary schooling was associated with lower fertility relative to women with no schooling. They were, however, unable to control for income. Shapiro and Tambashe (1994) found similar results in Kinshasa, Zaire, while controlling for "economic status." Fairlamb and Nieuwoudt (1991) found a large and significant negative effect of the years of female schooling on children ever born, but the results were confounded by endogenous regressors, and the study had no control for exogenous income. Ahn and Shariff (1994) found that seven or more years of schooling reduce the progression to first birth in Togo and Uganlda. Husband's education had a positive relation with cumulative fertility in four studies that did not have any other measures of income (Okojie 1990, 1991; Snyder 1974; United Nations 1987), but had a negative relation when womeni's education and a proxy for household permanent income were controlled for (Ainsworth 1989). A United Nations (1987) study of twelve African countries in the World Fertility Survey 90 THE. WORLD BANK ECONO(MIC REViFW. VOL. 10, NO. I found a negative relation between schooling and fertility for women with ten or more years of schooling and who had been married for at least three years. For levels of schooling less than ten years, the relationship was cited as positive. However, the statistical significance of the coefficients on lower years is not reported and female schooling was significant in only five to six of the countries. Husband's schooling at all levels had a positive relation with fertility, but was significant in only six of the countries. Multivariate studies of Sub-Saharan Africa at the individual level have shown a consistently positive relation between female schooling and contraceptive use in Sub-Saharan Africa.2 In comparing the determinants of contraceptive use over time using data from the 1977-78 Kenya Fertility Survey and the 1989 Kenya Demographic and Health Survey, Njogu (1991) found that educated women at all levels are significantly more likely to use contraception compared with women with no education, and that the impact of education rises with its level. An unpublished study of contraceptive use in Ouagadougou, Burkina Faso, showed similar positive effects of women's literacy and schooling on ever use of modern contraception (Fenn, McGinn, and Charbit 1987). All levels of schooling, in- cluding primary schooling, significantly raised the probability of contraception among a sample of nonpregnant women in Kinshasa, Zaire (Shapiro and Tambashe 1994). Similar results were found for a sample of all women of re- productive age in Burundi (Burafuta and Shapiro 1992). Castro Martin's (1995) results indicated, even at low levels of female schooling, statistically significant effects on contraceptive use among currently married women in ten African countries. Beegle (1995), Feyisetan and Ainsworth (1994), Oliver (1995), and Thomas and Maluccio (1995) found strong effects of female schooling on con- traceptive use when controlling for the availability (in Tanzania), quality (in Nigeria), and price of family planning services (in Ghana and Zimbabwe). In multivariate studies in other regions, women's education has a uniformly direct relation with contraceptive use and husband's education also has a direct but less powerful relation (Castro Martin 1995; Cochrane 1979). These studies suggest a negative but nonlinear relation between women's school- ing and fertility and demonstrate the potentially confounding effects of other vari- ables correlated with schooling (such as household income) in interpreting the re- sutts. However, many of the studies included potentially endogenous regressors in the analysis of fertility, variables such as the age at first birth, child spacing inter- vals, child schooling, and women's labor force participation (Snyder 1974); con1- traceptive use, breastfeeding, child survival, and the desired level of child schooling (Anker and Knowles 1982); and age at marriage (United Nations 1987). In multi- variate analysis of contraceptive use, endogenous regressors have included fertility, 2. The National Research Council (I 993) study conducted a multivariate analysis of the determinants of contraceptive use in sixty-eight regions of Sub-Saharan Africa using as regressors meani completed female schooling, urban residence, the percentage of women in polygamous unionis, and the percentage of women who practiced MLuslinm and traditional religions. The results confirned the very strong effect of female schooling relative to other factors. Ainsworth, Beegle, and Nyamete 9 1 women's labor force participation, current enrollment status (Burafuta and Shapiro 1992; Castro Martin 1995; and Shapiro and Tambashe 1994); marital status (Fairlamb and Nieuwoudt 1991; Fenn, McGinn, and Charbit 1987; Shapiro and Tambashe 1994); and fertility intentions (Fairlamb and Nieuwoudt 1991; Njogu 1991; Shapiro and Tambashe 1994). Decisions concerning marriage, timing of first births, child schooling, and labor force participation are jointly determined with fertility. Their inclusion will lead to biased results. An additional problem is that most of the studies used samples of currently married women. By conditioning on marriage, they have not captured the full effect of socioeconomic variables on fertility (through delayed marriage) or con- traceptive use (before marriage). Further, since marriage and childbearing could be thought of as a joint decision, studies that use married samples have introduced a potential sample selection bias in favor of women with higher demand for children. Snyder (1974) and Chernichovsky (1985) studied an even more selec- tive sample-women with clildren. Shapiro and Tambashe (1994) studied con- traceptive use in a sample of nonpregnant women and Castro Martin (1995) in a sample of currently married nonpregnant women. United Nations (1987) stud- ied marital fertility in a group of women married for at least three years. We seek to avoid manv of these pitfalls by using a common set of exogenous regressors on samples of all women, regardless of marital status. Further, by examining similar specifications, results can be compared across coLintries using the most up-to-date data. II. EMPIRICAL MODEL AND DATA SETS The empirical model of fertility determinants regresses a measure of cumula- tive fertility-children ever born to each woman-on a set of independent vari- ables that are assumed to be exogenous to fertility decisions but that influence either the demand for or supply of children. This reduced-form model of fertility determinants can be written as Y = A) + Aix1 + 312X + 2x+ /33x + 04X4+ fX + PbX6 + P7X7+ where y is the dependent variable, children ever born; P(, is an intercept; xl is the woman's age, entered in quadratic form to control for biological factors affect- ing the supply of births; x, is the woman's schooling, in various specifications; X3 indicates urban residence; x4 is the woman's ethnic group; x5 is the woman's region of residence; x 6 is the woman's religion; and x is a group of variables proxying the household's income or wealth. The empirical models for contra- ceptive use are identical, except for the dependent variable, which is a dichoto- mous (dummy) variable that takes on the value zero or one, indicating current use of modern contraception. To examine in more detail the nonlinearities of schooling effects and fertility, the results from four different specifications of female schooling are presented: (a) years of completed schooling as a linear term; (b) years of completed school- 92 THE WORLD BANK ECON(lMI(C REVIEW, VOL. 1W. NO. I ing in quadratic form (schooling and schooling squared); (c) linear splines for the level of schooling reached-lower primary (one to three years), upper pri- mary (four to six years or four to seven), lower secondary (seven to ten years or eight to ten) and higher levels (eleven years and more); and (d) dummy variables for individual years of completed schooling.3 Relationships for the national sample as well as for urban and rural subsamples are presented here. Results by cohort are presented in Ainsworth, Beegle, and Nyamete (1995). The choice of independent regressors has been influenced by the availability of information in the data sets. The analysis uses data from fourteen Sub- Saharan African countries collected by the DHSs in the late 1980s and early 1990s. The DHS samples are large-from 3,000 to 9,000 women aged fifteen to forty-nine. The large size of the samples is important because in many countries only a small proportion of women have had any schooling (fewer than a third of women in Burundi, Mali, and Senegal, for example). There are, unfortunately, no measures of household income or consumption in the DHS data sets. We use instead four variables that serve as proxies for income and wealth: household ownership of a radio, a television, or a bicycle, and the quality of housing. The latter is a dichotomous variable that indicates whether the floor of the household's residence is cement, tile, polished wood, or parquet; the default categories are mud, sand, or clay in most counitries. The number of children ever born is censored from below at zero, and can take on only zero or positive integer values. Under these conditions, least squares regression coefficients are inconsistent and, when the dependent variable and regressors are normally distributed, are biased downward in proportion to the degree of censoring in the sample (Greene 1981). Depending on the data set, roughly a quarter of the women in the samples (ranging from 16 percent in Mali to 30 percent in Burundi) have had no live births. Although there are econometric models capable of dealing with this problem, we have opted to use the easier-to- interpret ordinary least squares (OLS) estimation for the fertility regressions.4 This censoring problem does not arise for the contraceptive use regressions, in which the dependent variable is dichotomous. Maximum likelihood logit is used in estimation of the parameters (Maddala 1983).i 3. In six countries the prialary cycle is completed in sevtn years (Botswana, Kenya, Tanzania, Uganda, Zambia, and Zimbabwe). In these instanices the tipper primary and lower secondary variables are set equal to four to seveni and eight to ten years, respectively. Camerooni has a dual system, with those in the Francophonle areas completing primary schooling in six years, and those in the Anglophone areas in seven years. The tipper primary school spline Llsed in Caimeroon is set to reflect incomplete or complete primary schooling for the system in wlhichi the wyomana was enrolled. 4. The Tobit model, for example, takes into account the censoring of the dependent variahle at zero, although iot its integer nature A Poisson1 count model takes into account hoth the censoring and integer nature of the depetident variable, and the coefficients Would he Linbiased, even if the mean and variance of the dependenit vaLriable are not equal (as is implied ly the Poisson model). For all subsamples except younlg womien and possiblv urban woomen, Tobit and Poissoin models produce results identical in sign and significance and very similar in magnitude to ot s coefficients (Ainsworth 1989). 5. Standard errors in both the fertility and contraceptive tise regressions are corrected for heteroskedasticity anid cluster effects. A,nswortb, Beegle, and Nyamnete 93 A second censoring issue involves the inclusion of women who have not com- pleted their fertility. One way of dealing with this problem would be to predict the completed fertility of younger women on the basis of the determinants of fertility found among women who had already completed or nearly completed childbearing. However, in most African countries this would restrict the analy- sis to an older sample of women (over forty) that has had very little schooling. Further, we anticipate that the relation between schooling (as well as other in- dependent variables) and fertility may be changing over time. As the aggregate proportion of schooled women increases, the quality and price of schooling change, family planning becomes more widely available to younger women, and other socioeconomic conditions (such as income) evolve. To account for the fact that women in the samples have been exposed to the risk of pregnancy for dif- fering amounts of time and for the fact that fecundity over the fifteen-to-forty- nine age range rises with age, peaks, and then declines, we control for the woman's age and age squared. A third censoring issue of unknown magnitude, relevant to both the fertility and contraceptive use regressions, is the possibility that some of the younger women in the sample may not have completed their schooling. The DHS data sets did not record whether the women in the sample were currently enrolled in school. However, given the limited amount of schooling in all of the national samples at the primary and especially the secondary level, we do not anticipate that upper censoring of the completed schooling variable is a major problem in this analysis. 111. DESCRIPTIVF. STATISTICS The fourteen Sub-Saharan African countries included in this study span the continent in terms of their geographic location, colonial heritage, and level of economic development. Half of the countries are West African, four are in Eastern and Central Africa, and three are in Southern Africa. Five countries are French- speaking (Burundi, Mali, Niger, Senegal, Togo); eight are English-speaking, and Cameroon uses both as official languages. The countries range in size from Botswana, with slightly more than one million people, to Nigeria, the largest country in Sub-Saharan Africa, with a population of nearly 100 million (table 2). Most of these countries have very low incomes; as shown in table 2, seven had incomes per capita of less than US$400 in 1991. Among the twelve couIn- tries for which data are available, eight had negative rates of growth of gross national product (GNP) per capita in the 1980s, and for two others GNP per capita grew at less than 1 percent a year. Scribner (1995) has characterized the popula- tion policies of most of the countries in this study, including policies on family planning, child health, and women's legal status. There are wide divergences in female enrollment and infant mortality rates. As recently as 1990, for example, female primary enrollments represented only 17 percent of primary-age girls in Mali, 21 percent in Niger, and only Table 2. Economic and Social Indicators, Fourteen Sub-Saharan African Countries Grou'th rate of GNP per Population Female Female capita, growth rate, gross gross GNP per 1980-91 1980-91 Infant prinzary secondary capita, (average Population, (average Urban mortalitv enrollrzent enrollment 1991 (U.S. annual mid-1991 annual population rate, ratio, ratio, Country dollars) percent) (millions) percent) (percent) 1991 1990 1990 Tanzania 100 -0.8 25.2 3.0 34 115 63 4 Uganda 170 - 16.9 2.5 11 118 63b - Burundi 210 1.3 5.7 2.9 6 107 64 4 Mali 280 -0.1 8.7 2.6 20 161 17 4 Niger 300 -4.1 7.9 3.3 20 126 21 4 a1 Nigeria 340 -2.3 99.0 3.0 36 85 63 17 Kenya 340 0.3 25.0 3.8 24 67 92 19 Ghana 400 -0.3 15.3 3.2 33 83 67 31 Togo 410 -1.3 3.8 3.4 26 87 80 10 Zambia 420' - 8.3 3.6 51 106 91 14 Zimbabwe 650 -0.2 10.1 3.4 28 48 116 46 Senegal 720 0.1 7.6 3.0 39 81 49 11 Cameroon 850 -1.0 11.9 2.8 42 64 93 21 Botswana 2,530 5.6 1.3 3.5 29 36 112 47 - Not available. Note: Countries are listed in order of 1991 GNP per capita, from lowest to highest. a. Per 1,000 live births. h. 1992 data. c. 1990 data. Source: World Bank (1992, 1993, 1995). Ainsworth, Beegle, and Nyamete 95 half of the primary-age girls in Senegal (table 2). By comparison, the inclu- sion of over-age girls in primary enrollments in Botswana and Zimbabwe raises their gross female primary enrollment rates to over 100 percent. In only three countries-Botswana, Ghana, and Zimbabwe-do female second- ary enrollments equal or exceed 30 percent of the women in that age group. Perhaps not coincidentally, demographers believe that fertility decline is un- der way in two of these three countries-Botswana and Zimbabwe (van de Walle and Foster 1990). Infant mortality rates (IMRS) also show great varia- tion. Botswana and Zimbabwe have IMRs lower than 50 per 1,000, but six countries have rates greater than 100 per 1,000 (table 2). Despite these differences in economic and social indicators, levels of fertility and popula- tion growth rates in the fourteen countries are high and remarkably similar. With the exception of four countries, population growth rates were between 3 and 4 percent per year during the 1980s. Recall from table I that the total fertility rates were between 5.0 and 7.4 children per woman. Descriptive statistics for the fourteen DHS data sets and the year of data collec- tion are presented in table 3. Only four of the data sets are nationally self- weighted samples that achieved total coverage (Ghana, Senegal, Togo, and Zim- babwe). Thus, the unweighted sample statistics for other countries may not be nationally representative. A description of the sample design and coverage of the fourteen surveys is in appendix table A-i. The average woman in the samples in table 3 was twenty-seven to twenty-nine years old and, with the exception of Botswana, Ghana, and Zimbabwe, had three or more live births. Except for Botswana, Kenya, and Zimbabwe, fewer than 10 percent of the sampled women were using a modern method of contraception.6 Levels of urbanization in the samples were highest in Botswana, Cameroon, Mali, Nigeria, Senegal, and Zam- bia, where 40 percent or more of sampled women were living in urban areas. In Botswana, Cameroon, Mali, Niger, and Nigeria, urban women were oversampled. National levels of urbanization are reported in table 2. Women of reproductive age in these samples had very little schooling. The highest levels of schooling are for women in Zimbabwe, where the average woman in the sample had completed six years of schooling and where only 13.6 percent of the sample had no formal schooling. In Burundi, Mali, Niger, and Senegal, levels of schooling are lower: mean schooling was less than two years, and roughly three-quarters or more of the women had received no schooling. Since the DHS oversampled urban women in Burundi, Mali, and Niger, these statistics over- state the true levels of education at the national level. Figure 1 shows the distribution of women according to their years of com- pleted schooling. Countries are ordered from lowest to highest 1991 GNP per capita. Even in countries with relatively higher per capita incomes, like Senegal, a very large share of women have had no schooling. In fact, the French-speaking 6. Modern methods inclide female sterilization. vasectomy, pill, intrauterine device (IUD), injectable contraceptives, conidoms, spermicides, and diaphragm. Table 3. Demiiographic anzd Health Survey Samnple Statistics, Fouirteeni Sub-Saharaut Africani Countries Wom7aen Wf1sWomen Mean7 using modern living in Mean iuMber of contraCeptlu es urban x'ears of .Sample children (percent) areas teniale Distribution of wuomen by years of schooling (percent) Country Year size ever borm (Currenth' Euer (percent) scbooling 0 1-3 4-6 7-10 Ii or more Tanzania 1991-92 9,001 3.15 5.2 12.5 20.0 4.16 35.2 5.6 53.4 2.4b 34 Uganda 1988-89 4,727 3.40 3.4 9.6 20.4 3.67 34.5 17.0 35.8' 7.51 5.3 Burundi 1987 3,943 3.00 2.3 4.2 15.9 1.37 74.2 7.2 13.4 3.7 1.6 Mali 1987 .3,188 3.83 2.0 5.1 41.9 1.22 80.5 4.0 8.0 5.9 1.7 Niger 1992 5,854 3.78 4.6 8.8 38.8 1.04 83.1 2.3 8.1 5.3 1.1 Nigeria 1990 8,734 3.20 5.2 12.1 4(0.2 3.72 51.7 4.2 19.9 10.5 13.7 Kenya 1993 7,490 3.17 24.7 43.5 15.3 5.91 17.3 7.2 42.4' 20.6" 12.6 Ghana 1993 4,562 2.91 9.3 29.5 37.7 5.34 35.4 5.9 11.6 40.0 7.2 Togo 1989 3,351 3.21 3.4 10.7 35.2 2.31 58.8 10.1 19.1 1(0.4 1.6 Zambia 1992 7,004 3.13 6.7 22.1 47.6 5.37 18.4 11.3 48.4' 16.4' 5.4 Zimbabwe 1988-89 4,179 2.96 27.2 48.4 33.3 6.04 13.6 10.5 45.4a 16.6" 1.3.9 Senegal 1992-93 6,322 3.31 4.5 6.2 41.7 1.80 73.4 3.3 13.7 6.6 3.1 Cameroon 1991 3.862 .3.00 5.4 19.9 56.6 5.07 33.4 6.5 22.9' 24.5c 12.7 Botswana 1988 4,356 2.44 31.8 56.8 51.7 5.85 20.8 7.0 43.3d 21.6" 7.3 Note: Countries are listed in order of 1991 GNP per capita, from lowest to highest (see table 2). Statistics are for the unweighted working samilples used in the anialysis. See appendix rable A-I for samplinig strategies in each country. a. Includes four to seven years of primary schooling. b. Includes eight to ten years of schooling. c. In Cameroon, part of the country adheres to a six-year primary school cycle, and another part adheres to a seven-year cycle. The four-to-six category includes those who attended four to seven years in the seven-year cycle, and the secondary category represents seven to ten years or eight to ten years of schooling, depending on which system the woman attended. Source: DHS dara. Ainsworth, Beegle, and Nya7nete 9 7 Figure 1. Distribution qf Women 1 5-49 by Years of Completed Sckooling Tanzania, 1991-92 Uganda, 1988-89 Percent Percent 90 - 90 1 80- 80- 70 - 60 - 60-o 601 50 - 50 - 40- 404 20 20 10 10 0 0 . . . . . . 0 2 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 Years of schooling Years of schooling BrNdige, 1987Ngri,19 Percent Percent Mali, 1987 90_ 90 - 80 -80- 70 -70 -* 60 -60 -1 50 -50- 30 - ~~~~~~~~~~~30- 20 -20 10 10 I to 02 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 Years of schoolinlg Years of schooling (iueentigueria 1990efllu ag. Percent Niger, 1992) Percent ie-i,19 90 90 80 80- 70 - 70- 6011- 60- 50 -1 50- 30 -30- 20 20- 10 1110( 0246 87 10 12 14 16 18 20 2 2 0 24 10 12 14 16 18 20 )2 Years of schiooling Years of schooling Kenya, 1993 Glhana, 1993 Percent Percent 90 901 80 -80~ 70 -O7 60 -60 - 50 -5 40 0 30 -301f 20 2011 10 __ _ _ _ ___0_ __ _ _ _ 0 0 11 1 11 1 1 0 2 46 81012 14 1618210212 02 46 811012 1416 1820 22 Years of schloolinig Years of schlooling (Figure continues on the following page.) 98 THE WORLD BANK ECONOMIC REVIEW, VOL. 10, NO. I Figure 1. (continued) Togo, 1989 Zambia, 1992 Percent Percent 90 - 90 - 70- 70- 60- 1 60- 501 501- 40 40 - 30j- 30- 20 - ^g20 - o ;. 0=W , , '|| lo G l ' , 02 4 6 8 10 12 14 16 18 20 22 0 2 4 6 8 10 12 14 16 18 20 22 Years of schooling Years of schooling Zimbabwe, 1988-89 Senegal, 1992-93 Percent Percent 90- 90 - 801 80 70 - 70 - 60 60 - 505- 50 - 40 - AO40 30 30- 20 20 0 2 4 6 8 10 12 14 16 18 20 22 024 6 8 10 12 14 16 18 20 22 Years of schooling Years of schooling Cameroon, 1991 Botswana, 1988 Percent Percent 90-F 90 - 801- 80 - 701 -70 601- 60 - 50 - 50 - 40 -40 - 30- 30- 20 Io2-u 10 4 0-4 ~- 0 46810121416 1,82,022 0 2 4 6810 121 1416 1820 22 Years of schooling Years of schiooling Note: CoLuntries are in order of 1991 GNP' per, capita, frorm lowest to highest (see table 2). For sample size for each couintry, see table 3. Souirce: DI) is data. Ainsworth, Beegle. and Nvamete 99 and Sahelian countries have by far the lowest schooling of women of reproduc- tive age. But even in Ghana and Nigeria, thought to have strong educational policies in the past, 35 and 50 percent of all women in these samples, respec- tively, have had no schooling. In Tanzania, once thought to have achieved uni- versal primary education, about 35 percent of women never attended school. Of course, figure 1 groups together women of many different birth cohorts. When the distribution of schooling is compared across cohorts, substantial progress has been made in raising female schooling over time in Botswana, Cameroon, Ghana, Kenya, Tanzania, Zimbabwe, and Zambia. However, in Burundi, Mali, Niger, and Senegal there has been little or no change in the level of female education over time.' The distribution of schooling among educated women is often uneven. In Ghana, for example, about 35 percent of women had no schooling, and about 25 percent had exactly ten years. In Botswana, Kenya, Nigeria, and Zimbabwe, "spikes" are observed in the distribution at seven years of schooling. The lumpi- ness of the distribution of schooling in some of these data sets-including the relative scarcity of observations with less than complete primary schooling-is important to consider in interpreting the regression results. While the mean level of schooling may be two to three years, very few women are commonly found with exactly that number. IV. WOMEN'S SCHOOLING AND FERTILITY The OLS regression results for three of the four specifications in the fourteen countries are pictured in figure 2. On the basis of the regression coefficients, predicted children ever born is plotted against women's years of schooling, while controlling for age, age squared, area of residence, ethnicity, religion, owner- ship of durable goods, and quality of housing. The straight line in each graph represents the linear specification of female schooling, in which the slope is con- strained to be constant. In all of the countries, the linear specification is sloped downward, showing that increased schooling is generally associated with lower fertility. However, the descriptive statistics suggest that the relation between female schooling and fertility is nonlinear. The broken lines in figure 2 represent the predictions from a quadratic specification of female schooling (schooling and schooling squared), and the small circles represent the prediction of a specifica- tion allowing a dummy variable for each individual year of schooling. These two specifications show that at low levels of schooling the relation between female schooling and fertility is weak or nonexistent, but that with the comple- tion of primary schooling the relation is clearly negative. This makes sense, since it is difficult to believe that less-than-full primary schooling could result in basic 7. For the distribution of schooling in urban and rtural areas and by cohort, refer to annex 2 of Ainsworth, Beegle, and Nyamnete (1995). 1 00 THE WORLD BANK ECONOMIC REVIEW, VOL. 10, NO. I Figure 2. Predicted Cumulative Fertility by Female Schooling Using Three Specifications Tanzania Uganda Children ever bom Children ever bom 5 ~~~~~~~~~~5- 4- 4 2 2 0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16 Years of schooling Years of schooling Burundi Mali Children ever bom Children ever bom 5~ 5_ 4- 4 ..._ Q 2 2 oo ; ,, > o2 : 0 0~~~~~~~~~~~~~~~~~~~~~~ O- 1 l*- l l l0 0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16 Years of schooling Years of schooling Niger Nigeria Children ever bom Children ever bom . 5~~==5~EC 5- _______________________ &~~4 a. 00 2 2 0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16 Years of schooling Years of schooling Kenya Ghana Children ever bom Children ever bom 4 4 40 2%1 4 62 14 16 0 2 4 6 81 12 14 16 Years of schooling Years of schooling Ainswvortb, Beegle, and Nyamnete 101 Figure 2. (continued) Togo Zambia Children ever born Children ever born 5- 9 2 2 l 1 0 0 2 4 6 8 10 12 li 16 0 2 4 6 8 10 12 14 16 Years of schooling Years of schooling ZimbabvweSega Children ever born Children ever born Senegal 1 1 I 0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16 Years of schooling Years of schooling Cameroon Botswana Children ever born Children ever born '4 -j 04 2 j I S 11 0 0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16 Years of sclooling Years of schooling iVote The graphs show predicted number of children ever born from ordinary least squares regressions ulsing three different specifications of female schooling: the straight line in each graph represents a linear specification; the broken line represents the predictions fromil a quadratic specification (schooling and schooling squared); and the circles represent the prediction of a specification that allows a dummy variable for each individtial year of schooling. In the latter, solid circles represent statistically significant coefficients (at the 5 percent level or less) Countries are in order of 1991 GNP per capita, from lowest to highest (see table 2). For sample size for each country, see table 3- Source Authors' calculations from DHS data. 1 02 THE WORLD BANK ECONOMIC REVIEW. VOL. 10. NO. I Table 4. The Impact of Women's Schooling on the Number of Children Ever Born, Fourteen Sub-Saharan African Countries Level of schooling (years) Country Years of schooling 1-3 4-6 7-10 11 or more Tanzania -0.075*' -0.124 -0.324` -0.255 -1.201* Uganda -0.060#* -0.02-7 -0.022 -0.271 -1.419'> Burundi -0.024 0.182 0.141 -0.366*# -0.823** Mali -0.069* * -0.009 -0.027 -0.516- -1.483* " Niger -0.060' -0.269 -0.280* -0.736 # -0.178 Nigeria -0.t34` 0.319^* -0.2981- -0.722**'t -1.626"# Kenya -0.126'i 0.173 -0.257* -0.671 ` - 1.464* Ghana -0.0 9*' 0.094 -0.104 -0.447" -1.4312 Togo -0.0i61 -0.005 -0.0.5 -0.483-` -1.209#* Zambia -0.092:- 0.207* -0.045 -0.546 * -1.S20** Zimbabwe -0.102' -0.022 -0.36.5 -0.646* -1 .184* Senegal -0.I02# -0.002 -0.283'- -0.667'* -1.816** Cameroon -0.116 --0.(05 -0.174 -0.465:" -1.412"" Botswana -0.086 ;" 0.068 -0.222"` -0.521' ' -1.245** Significant at 5 percent level. ** Significant at I percenit level. Note: Coefficients were estimated using ordinary least squlares (olis). For years of schooling, the specification was linear; for level of schooling, the specification ised dummy variables to estimate the effect on the number of children born of raising female schlooling from zero years to each level. Countries are listed in order of 1991 (INI' per capita, from i lowest to highest (see table 2). Source: Authors' calculations, I)Hs data. literacy and numeracy or could substantially alter the opportunity costs of women's time. Also recall that most women are in the "no schooling" category or in the complete primary schooling category. Thus, some of the dummy vari- ables for individual years of schooling at both incomplete primary and second- ary levels represent very, few women and are not individually statistically signifi- cant. However, a few outliers can greatly distort the relationship, as in the case of Niger. In fact, given how few women obtained higher levels of schooling in the Sahelian countries, the curves for very high schooling levels must be viewed with skepticism. Table 4 shows regression coefficients for the linear specification and for a specification not shown in figure 2, which used dummy variables representing four different levels of schooling-one to three years, four to six years, seven to ten years, and eleven or more years.8 The coefficients are interpreted as the effect on the number of children ever born of raising female schooling from zero to that level. All other explanatory variables (age, age squared, urban residence, region, ethnicity, religion, ownership of durable goods, and quality of housing) were also entered in the regressions when available.9 8. For counitries with a seven-year primary cvcle, the levels entered are one to three, four to seven, and eight to ten years. 9. The following data sets did niot have the full set of indepenidenit variables: ethnic group was missing for Botswana, Burunidi, Cameroon, Nigeria, Tanzaniia, Uganda, and Zambia; regioni was missing for Botswana; and religion was missing for Buiuldi and Senegal. Ainsworth, Beegle, and Nyanmete 103 The number of years of female schooling is significant and negatively related to cumulative fertility in thirteen of the countries, despite their different levels of female schooling and economic characteristics. With the exception of Senegal, the largest linear female schooling coefficients occur in countries in which the samples have large shares of women with postprimary schooling (Cameroon, Kenya, Nigeria, and Zimbabwe; refer to table 3). However, Ghana has the larg- est share of sample women with postprimary education (47 percent; see the last two columns in table 3) and the linear relation there is relatively small. The results for the levels-of-schooling specification in table 4 indicate that lower primary schooling (one to three years) is not related to cumulative fertil- ity in twelve countries and has a positive relation in two (Nigeria and Zambia). In half of the countries, women with four to six years of primary schooling have 0.2 to 0.4 fewer children ever born, compared to women with no schooling, and in the other half there is no relationship. On average, controlling for covariates, women in the samples with seven to ten years of schooling had from 0.2 to 0.7 fewer children ever born, and women with eleven years of schooling or more had 0.8 to 1.8 fewer children ever born, compared with women with no schooling. Much has been made in the literature of the sometimes-observed positive relation between a few years of schooling and fertility. This is often explained as inadvertent outcomes of changes in proximate determinants of fertility. A posi- tive relation would imply that even one year of primary schooling is sufficient to induce a quite large change in behaviors and outlooks that indirectly affect fertil- ity. The realities of primary schooling in most Sub-Saharan countries-in terms of poor infrastructure, extremely limited availability of reading and instructional materials, inadequate teacher training and salaries, and the resulting high ab- senteeism and dropout rates-cast doubt on this interpretation (Lockheed and Verspoor 1991). Literacy is certainly not achieved in a single year of schooling and, under the circumstances that prevail, may not even be achieved until comple- tion of primary school. A more plausible explanation is that the (small group of) women who completed only a few years of schooling are those who became pregnant, whose families wanted them to get married, or who simply could not keep up and therefore stopped their schooling. Another plausible explanation for the sometime nonrelation between primary schooling and fertility or a positive relationship is the exclusion of variables like household income from the regressions. If, as incomes rise, parents want more children, holding wages and other prices constant, then higher incomes should be associated with higher fertility. If there is a strong association between schooled women and higher incomes, and if income is not properly controlled for, then the schooling coefficients may be absorbing both the negative schooling and positive income effects, which cancel out. This would also implv that the co- efficients on women's schooling are underestimates of the negative relation with fertility. This was found to be the case in C6te d'lvoire, where omitting income from fertility regressions notably weakened the schooling coefficient in magni- 104 THE WORLD) BANK ECONOMIC REVIEW, VOL. 10, NO. I tude (Ainsworth 1989). The controls for household assets-durable goods and quality of housing-may not have completely alleviated this problem in this study. By the later years of primary schooling, however, some degree of literacy should have been achieved, even for women who did not complete primary school. Thus, we expect a negative relation in these cases. However, the last years of primary schooling are associated with lower fertility in only half of the countries and have no association in the other half. Understanding why this relation exists in some countries and not others is important to policymaking. It might reflect differences in the quality of instruction or differences in labor market conditions across countries, for example. Part of the explanation may also lie with complementaritv between schooling and the availability of family planning ser- vices; Botswana, Kenya, and Zimbabwe have the strongest family planning pro- grams and also show negative relations between female primary schooling and fertility. The strong negative association between women's higher secondary school- ing and cumulative fertility evidenced in figure 2 and table 4 persists across all of the countries in the sample, regardless of their level of development, distribu- tion of schooling, or intensity of family planning programs. (Niger is an excep- tion; fewer than 0.1 percent of women had eleven or more years of schooling.) These results suggest that, with or without easy access to family planning, highly educated women do manage to lower their fertility. When supplies are scarce, highly educated women have the greatest access to contraceptive services by virtue of their education and probable income levels. However, in three coun- tries with more active family planning programs, even women with only pri- mary instruction are able to lower their fertility. There is no obvious relation between the size of the linear schooling coefficients in table 4 and GNP per capita. This is not surprising, because cumulative fertility is likely to reflect past levels of income. Nigeria's income levels, for example, are vastly lower now than they were during the oil boom of the 1970s. In addi- tion, GNP per capita hides differences in the distribution of income that may account for differential effects of schooling. Within countries, the regression results for ownership of assets revealed a negative relation between income and fertility in Botswana, Ghana, Tanzania, and Togo, but a positive relation in Nigeria. Results for other countries were ambiguous (see Ainsworth, Beegle, and Nyamete 1995). Urban and Rural Samples In the Botswana, Cameroon, Mali, Nigeria, Senegal, and Zambia data sets, 40 percent or more of women were living in urban areas. The least urbanized samples are in Burundi, Kenya, Tanzania, and Uganda, with 20 percent or fewer women in urban areas. Figure 3 shows predicted cumulative fertility using a linear spline specification of female schooling at zero to three, four to six , seven to ten, and eleven or more years of schooling, in urban and rural areas (Greene 1993). The Ainsworth, Beegle, and Nvanmete 1 05 Figure 3. Predicted Cumtulative FertilitV by Female Schooling in Uirban and Rural Areas Tanzania Uganda Children ever born Children ever born 5 ~~~~~~~~~~~~~~~~5 4 4 3 3 22 1 1 0 2 4 6 8 101214 16 0 2 4 6 810 121416 Years of schooling Years of schooling Burundi Mali Children ever born Children ever born 4 3~~~~~~~~~~~~~~~~~~ 2 1 0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16 Years of schooling Years of schooling Niger Nigeria Children ever born Children ever born 5 5 33. 2 2 1 1 0 0 0 2 4 6 8 10 12 14 16 ( 2 4 6 8 10 12 14 16 Years of schooling Years of schooling Kenya Ghana Children ever born Children ever borm 5 - ~~~~~~~~~~~~~~~~5 3 3 2 2 0 o~~~~~~~~~~~~~~~ 0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16 Years of schooling Years of sclooling Urban - RLIral (Figt-c contnles on the follou ing page.) 106 THE WORLD BANK ECONOMIC REVIEW, VOL. 10, NO. I Figure 3. (continued) Togo Zambia Children ever born Children ever born 4 4 3 3 2 2 1 1 0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16 Years of schooling Years of schooling ZimbabweSega Children ever born Children ever born Senegal 5 ~~~~~~~~~~~~~~~~~5 4 4 - - 3- = = ----------------- 3- 3 3 2 .............2 1 1 0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16 Years of schooling Years of schooling Cameroon Botswana Children ever born Children ever born 5 5 4 4 3 3 22 0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16 Years of schooling Years of schooling Urban ------------- Rural Note. The graphs show predicted cumulative fertility using a linear spline specification of female schooling with segments for 0-3, 4-6, 7-10, and 11 or more years of schooling. Countries are in order of 1991 GNP per capita, from lowest to highest (see table 2). For sample size for each country, see table 3. Source: Authors' calculations from DHS data. Ainsworth, Beegle. and Nvanmete 107 linear spline specification is jointly significant for all countries in urban and rural areas, with the exception of urban areas in Niger and rural areas in Burundi. However, specific segments of many of these curves are not statistically signifi- cant. The coefficients underlying figure 3 can be obtained from the authors. In interpreting the results, it is important to keep in mind that 90 percent or more of rural women in Mali, Niger, and Senegal, and 64 percent of urban women in Mali and Niger, had no schooling. Thus, the results for higher-order splines are based on very few cases and the results are highly sensitive to outliers. With these caveats in mind, differences between urban and rural women within and across countries are nevertheless interesting. Fertility declines with increases in female schooling in both urban and rural areas, particularly after primary school- ing. The early years of primary schooling are associated with higher fertility in urban Nigeria and Uganda, but are otherwise insignificant. In most of the coun- tries, family planning services are not easily available in rural areas, yet very educated rural women (there are few of them) nevertheless have lower fertility. In fact, the differential between women with eleven or more years of schooling and those with no schooling is often greater in rural than urban areas (Cameroon, Kenya, Niger, Nigeria, Uganda, and Zambia) (see figure 3). In general, at every level of schooling urban women have lower fertility than rural women. This may be due to a variety of factors, including more labor market opportunities, higher costs of children, and more readily available health and family planning services in urban areas. Nigeria is an important excep- tion-at the highest levels of schooling, fertility declines more rapidly in rural areas and is in fact lower than in urban areas. Fertility in urban and rural areas also converges at higher levels of female schooling in Botswana, Kenya, and Uganda. When controlling for female schooling, differentials between urban and rural fertility are quite small in Botswana, Cameroon, Nigeria, and Zam- bia, but remain large in Ghana, Senegal, and Togo. Results by Cohort The relation between women's schooling and fertility may change over time. As female enrollment rates rise, a greater percentage of each successive cohort has had schooling. The level of schooling of the woman relative to the schooling of her cohort may alter the returns to schooling and thus the relation between schooling and fertility. Even if enrollment rates were to remain constant, changes in the quality and content of schooling over time might result in changes in the effectiveness of women's schooling in altering fertility behavior. Finally, over time other variables in the environment may also change, enhancing or detracting from the relationship. For example, making family planning more available might alter the effect of schooling on cumulative fertility by substituting for the schooling of poor women (lowering fertility even for women with no schooling) or by availing more educated women of the means to keep their fertility low. Regressions run by cohort (ages fifteen to twenty-four, twenty-five to thirty- four, and thirty-five to forty-nine) reveal that the overall relation between 108 THE WORLD BANK ECONOMIC REVIEW, VOI. 10, NO. I women's schooling and cumulative fertility is negative in the youngest and middle- age cohorts in all of the countries and in the oldest cohort in more than half of the countries (results are available from the authors). The negative relation be- tween schooling and cumulative fertility is almost always larger in the middle cohort than in the youngest cohort. The results suggest that the effect of women's schooling on fertility increases with the woman's age, and middle-aged edu- cated women are not "catching up" with the fertility of others in their cohort to compensate for lower fertility while young. The association of partial or com- pleted primary schooling with lower cumulative fertility is observed to a greater extent among the youngest cohort. However, Botswana, Ghana, and Zambia, which have relatively high levels of schooling, show no significant effect of pri- mary schooling in any of the cohorts. The higher levels of schooling are asso- ciated with significantly lower fertility in all cohorts, even in countries with low levels of schooling in the older cohort. Women's Schooling Comnpared with Men's An issue of policy interest is the relative impact of men's and women's schooling on cumulative fertility. Because child rearing is not generally intensive in the time of the father in these countries, we do not expect higher education and earning capacity of husbands to raise the "price" of children, as does women's education. More often, husband's education is used as a proxy for household income, and it may be a better proxy for income than the ownership of durable goods, which are already controlled for. Inclusion of an income proxy may alter the size of the woman's schooling coefficients to the extent that they are corre- lated with income or husband's education. However, to examine these issues we must restrict ourselves to the sample of ever-married women in each data set. Depending on the country, the sample of ever-married women may be as small as 45 percent of the total sample (Botswana) or as large as 86 percent (Mali). The restricted sample, conditioned on marital status, can be expected to show a smaller schooling effect for ever-married women than for all women, because the influence of schooling on delayed age at first birth of the unmarried women in the sample will not be included. Table 5 presents the OLS regression results of children ever born on the linear specification of woman's schooling, with and without the husband's schooling. The woman's schooling coefficients for the sample of ever-married women in table 5 (second column) are generally smaller than for the sample of all women in table 4, and in Uganda and Zimbabwe the difference is considerable. This result confirms the point made earlier that studies restricted to samples of ever- married women may understate the relation between female schooling and fertility. In table 5, controlling for husband's education generally acts to reduce slightly the woman's education coefficient (column three). In four countries (Cameroon, Kenya, Nigeria, and Uganda) only the woman's schooling is associated with lower fertility; the coefficients on husband's schooling (column four) are not significant. When statistically significant, increases in the husband's schooling are associated Ainsworth, Beegle. and Nvaimete 1 09 Table 5. The Impact of Women's and Their Husbands' Schooling on Cumulative Fertility, Fourteen Sub-Saharani African Countries Regression witbout Regressioni with Percentage with no husband's education husband's education schoolinig Coefficient on Coefficient on Coefficient on Ever- Sample woman 's years woman;'s years hzusband's vears married Country size of schooling of schooling of schoolinig women Husbahnds Tanzania 6,593 -0.069K* -0.063* -0.017* 42.5 27.7 Uganda 3,657 -0.034* -0.031 -0.006 40.1 18.0 Burundi 2,541 -0.015 -0.019 0.011 77.3 57.7 Mali 2,750 -0.062* * -0.045*' -0.026* 83.0 84.3 Niger 4,953 -0.007 0.012 -0.058X 87.6 89.2 Nigeria 6,912 -0.112- -0(.114 0.002 62.5 53.4 Kenya 5,003 -0.106 * * -0.09X i^ -0.015 23.1 13.0 Ghana 3,417 -0.078** -0.063- -0.047** 41.6 31.9 Togo 2,318 -0.054* * -0.042` -0.022* 68.0 49.7 Zambia 5,115 -0.088** -0.076(< -0.022- 21.2 11.1 Zimbabwe 2,75 5 -0.079* -0.(62) -0.038X 16.3 12.4 Senegal 4,354 -0.096^* -0.07 3 ' -0.034 i* 85.2 84.2 Cameroon 2,878 -0.11 1* -0.10w* -0.014 41.3 36.3 Botswana 1,954 -0.094* -0.(70' -0.044"- 27.9 35.4 Significant at the 5 percent level. Significant at the I percent level. Note: The samples include only women who have ever beeen married. Coefficients were estimated using ordinary least squares (o1 ,) and a linear specification of the vears of temale and male schooling. Countries are listed in order of 1991 GNP per capita, from lowest to highest (see table 2). Source: Authors' calculations based on l-HI data. with lower fertility, but the negative association between women's schooling and fertility is stronger than the men's in all but one of the remaining countries. Only in Niger is the husband's education alone associated with lower fertility. Investments in women's schooling, therefore, are likely to have a greater impact on fertility than investments in men's schooling-and the effects of women's schooling are even greater in the sample of all women than in the sample of ever-married women. V. WOMEN'S SC HOOLING AND CONTRACEPTIVE USF Contraceptive use is related directly to the demand for children. Therefore, all of the factors leading educated women to have fewer children should result in a positive relation between education and contraceptive use. In addition, edu- cated women may be more likely to use contraception because information about the availability, correct use, side effects, costs, and so forth may be less difficult and costly for educated women to assimilate, and may make them more effec- tive and satisfied users. Levels of current use of modern conitraception in the DHS samples of all women used in the regressions are presented in table 6, for urban and rural women and for different levels of schooling. Both in our data and in virtually all tabulated results to date, urbani women and women with more school- ing (even those with primary schooling) have higher rates of contraceptive use I I 0 FtILE WORLD BANK ECONOMIC REVIEW, VOL. 10, NO. I Table 6. Women Currently Using a Modern Method of Contraception, by Womiian's Residenice anid Education (percent) Woman's residence Woman's education (years) Country Year All Urbait Rural 0 1-6 7-10 11 ormore Tanzania' 1991-92 5.2 10.8 3.8 2.1 6.1 10.7 16.7 Uganda' 1988-89 3.4 10.7 1.5 1.0 2.8 9.6 15.6 Burundi 1987 2.3 11.5 0.5 0.8 4.9 9.7 19.4 Mali 1987 2.0 4.8 0.1 (.7 3.9 12.2 17.0 Niger 1992 4.6 1 0.7 3.3 10.0 13.5 15.2 Nigeria 199( 5.2 9.9 2.1 1.8 6.4 7.5 14.3 Keniva 1993 24.7 32.2 23.3 17.0 23.1 24.4 42.1 Ghana 1993 9.3 12.5 7.3 3.8 9.8 11.8 20.7 Togo 1988 3.4 6.2 1.9 1.8 4.4 8.3 14.8 Zambia, 1992 6.7 10.9 3.0 2.7 4.8 11.4 28.2 Zimbabwvel 1988 277.2 33.6 24.0 23.5 27.8 23.3 33.0 Senegal 1992-93 4.5 9.0 1.3 2.2 8.4 13.6 18.8 Cameroon 1991 5.4 7.8 2.3 1.7 4.6 7.3 13.0 Botswana, 1988 31.8 38.1 25.0 18.7 30.4 39.9 54.3 Note: Unfortunately, these differentials are only produced in the DHS countrv reports for ever-married women. Thus these differentials in contraceptive prevalence rates have been calculated from the data used for analysis and are not weighted to compensate for oversampling. Countries are listed in order of 1991 (,NP per capita, from lowest to highest (see table 2). a. Primary schooling is one to seven vears, and the next highest level is eight to ten years for these countries. Source: Authors' calculations based on DHS data. than rural and less-educated women (Castro Martin 1995; Jejeeboy 1992; Na- tional Research Council 1993). Figure 4 shows the predicted relation between years of women's schooling and contraceptive use in the sample of all women, using a spline specification and holding the value of all other variables in the regressions at their means. These predictions are based on the results of logit regressions of current or ever use of contraception on the same set of independent regressors as in the fertility work: age; age squared; urban residence; female schooling; dummy variables for ownership of a bicycle, a radio, or a television; dummy variable for the quality of flooring; religion; region; and ethnic group. In order to maintain cross-country comparability, controls for the availability of family planning are not included here. Availability of family planning was collected in four of the DHS data sets (Kenya, Nigeria, Tanzania, and Zimbabwe) and has been analyzed by Beegle (1995) for Tanzania, Feyisetan and Ainsworth (1994) for Nigeria, and Thomas and Maluccio (1995) for Zimbabwe. Using Living Standards Survey data, Oliver (1995) has analyzed the impact of the availability of family planning in Ghana. In all countries an increase in female education is significantly associated with an increase in current contraceptive use.10 However, in many countries- 10. In Niger and Zimbabwe, the splines are jointly significant for current use at p<0.05, even though none of the individual coefficients were significant. However, the splines in tihe ever-use regression were not significant for Niger. Ainsworth, Beegle, and Nyainete 1 17 Figure 4. Predicted Current Use and Ever Use of Molern Contraception by Female Schooling Tanzania lUganda Probability of contraceptive use Probability of contraceptive use .7 :7 .6 .6 .5 . .4 . .3 .3 .2 ~~~~~~~~~~~~~~~~2- .1 .1~~~~...................------ 0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16 Years of schooling Years of schooling Burundi Mali Probability of contraceptive use Probability of contraceptive use .8 22 .7 .7 .6 .6 .5 .4 . .3 . .2 .2 .1 .1 0 2 4 6 8 1012 14 16 0 2 4 6 8 1012 14 16 Years of schooling Years of schooling Niger Nigeria Probability of contraceptive uise Probability of contraceptive use .8 .8 .7 .7 .6 .6 .5 . .4 . .3 .3 .2 .2 .1 .1 - ........... 0 2 4 6 8 10 12 14 16 0 2 4 06 8 1-0 12 1416~ Years of schooling Years of schooling Kenya Ghana Probability of contraceptive use Probability of contraceptive use .7 . .6 .0 .5 .5 .4 - .3 ......3 - .2.. . . . . . . . . . . . 0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16 Years of schooling Years of schooling Ever use ------------Current use (Figure continues on the followXing page.) 1 2 TI.F WORLD BANK EC.oNOMIC REVIEW, VOL. 10, NO I Figure 4. (continiued) Togo Zambia Probability of contraceptive uIse Probability of contraceptive use 08 2 4 6 8 10 12 14 16 0 2 4 6 8 10 1214 168 -7 .~~~~~~~~~~~~~~~~~~~~~~~~~~~~~7 .6 .6 .5 .4 .4 .3 .3 .2 .1 .~~~~~~1 1 1 I I 1 0 2 4 6 8 10 12 14 16 0 2 4 6 8 1 2 14 .6 Years of schooling Years of schooling Zimbabwe Senegal Probability of contraceptive Use Probability of contraceptive use 8 O.8 .6 .6- .5 .4 . .1 .1 01 1 1 1 1 1 1 1 (. ~~~~~~~~~~~~~~~~~~~~~I I I I I I I O 2 4 6 8 10 2 14 1 0 2 4 6 8 1012 14 16 Years of schooling Years of schooling Ca meroon Bots-ana Probability of contraceptive ased Probabilitv of contraceptive use .8 . .6 .6 .3 .3 .2 . .... . .1 .. .. . . ... . . ... . . .. 0 2 4i 6 8 1t 12 114 16 0 2 4 6 8 10 12 14 16 Years of schiooling Years of schooling Ever uiSe--------Current Luse Note.- Thec predictionis are based on logit regressions of current Luse and ever uise of contraceptives, Lusing a spline specification (0-3, 4-6, 7-10, and 11 or inore years of schooling) and holding the values of all other variables in the regressions at their means. Countries are in order of 1991 GNP per capita, front lowest to highest (see table 2). For sample size for each countrv, see table 3. Source.- Authors' calculations from tD5s data. Aissworth, Beegle, and Nyamete 1 13 most with very low levels of female schooling and limited availability of family planning-both current and ever use are low, even among women with high levels of schooling. The major increase in current contraceptive use in Burundi, Ghana, Niger, Senegal, and Tanzania occurs at the early years of female school- ing, and there is very little increase in contraceptive use with additional school- ing beyond primary. In fact, even in countries where higher levels of female schooling show positive gradients (Botswana, Cameroon, Kenya, and Zambia), the slope of the relationship at earlier years of schooling is steeper. Only in Zimbabwe does the relationship between female schooling and current con- traceptive use seem to steepen with increased schooling levels. However, it is important to remember that there are very few women with high levels of school- ing in the majority of these countries. Thus, this last segment of the spline may be very sensitive to outliers and responsible for seemingly weaker (or contrary) results for higher levels of schooling in some of the countries. The education- contraceptive use gradient is steeper for ever use of contraception as education rises, with the exception of Niger (where there is no significant relation for ever use) and Senegal and Zimbabwe (where the relationship is basically flat for all but the early levels of schooling). The relationship between female schooling and contraceptive use does not seem to have any relation to the level of GNP per capita. For example, Senegal and Zimbabwe have the same GNP per capita, but the relationship is quite different in the two countries. Within countries, owner- ship of assets was associated with higher contraceptive use, particularly in Kenya, Nigeria, Tanzania, Zambia. and Zimbabwe (Ainsworth, Beegle, and Nyamete 1995). These results represent the relationship between female schooling and con- traceptive use at the time of the surveys, holding all other factors constant. However, they do not represent an immutable relationship or law unique and unchanging for every country. As the average level and quality of female school- ing and the returns to education rise in countries like Burundi, Mali, and Niger, if the experience of other countries is a guide, their curves will very likely shift upward. Likewise, the steepness and height of the relationship should also be related to the availability of contraception. In Zimbabwe, for example, the slope of the curve is fairly flat. It may be that the greater availability of contraception mutes differentials in contraceptive use by education in Zimbabwe relative to Kenya, which shows sharp curvature upward. Figures for Zimbabwe and Botswana thirty years ago would have looked quite different than they do now, following ma jor investments in female schooling and better contraceptive services. Urbani and Riural Samples The effect of schooling on contraceptive use is generally greater in urban than in rural areas (not shown). At the mean level of female schooling in the samples, an additional year of schooling is associated with 0.4 to 2.6 percentage points increase in current contraceptive use in urban areas and 0.1 to 1.8 percentage points increase in rural areas. In urban areas the early years of primary school- 14 THF WORLD BANK ECONOMI(. REVIEW. VOl.. lo. NO. I ing are associated with higher contraceptive use only in Mali. However, in nine of the countries the later years of primary schooling have an impact; among those for which this is not the case are countries with relatively high average schooling levels-Cameroon, Ghana, Zambia, and Zimbabwe. The only coun- try for which higher levels of schooling do not significantly affect contraceptive use in urban areas is Zimbabwe. Possibly the relatively wider availability of contraceptives in urban areas of that country mutes differentials in contracep- tive use by female schooling. Contraceptive use is so low in rural areas that regressions could be run on rural data from only eight countries. The early years of primary schooling are more likely to be associated with increased contracep- tive use in rural areas than in urban areas. Women's Schooling Compared U'ith Men's Table 7 classifies countries according to the relative strength of the relation- ship between the woman's and husband's schooling and current contraceptive use, among the sample of ever-married women. In half the countries, only the woman's schooling is a statistically significant determinant of current contracep- tive use. This group includes countries with both the highest and lowest average levels of female schooling, income, and availability of contraceptives (for ex- ample, Botswana and Zimbabwe, as well as Mali and Niger). In six countries, both the woman's and husband's schooling are significant. However, in all of these countries the coefficient on the woman's schooling is greater in absolute value than the husband's schooling coefficient. What can explain the unusual grouping of seemingly different countries? The countries for which only female schooling is a statistically significant determi- nant of contraceptive use are those for which husbands and wives have roughly equal probabilities of having received instruction. In Mali and Niger, for ex- ample, from table 5, the percentage of wives and husbands without any school- ing is high and roughly equal, while in Cameroon and Zimbabwe the percentage is lower but still similar between husbands and wives. In Botswana, fewer hiis- bands than wives have any schooling. The most important exception is Uganda, which has a spread of 22 percentage points between wives and husbands in the percent with any schooling. In the countries where both husband's and wife's schooling are significant determinants of contraceptive use, the differential be- tween the percentage of husbands and wives with any schooling is generally greater. An important exception is Burundi, where the differential between the percentage of husbands and wives with no schooling is great, but neither is signifi- cant. Another important exception is Senegal, where there is basically no difference between the percentage of husbands and wives with no schooling, but both the wife's and husband's schooling coefficients are significant. The observation that women and their husbands have similar schooling levels in countries where only the woman's schooling is a significant deter- minant of contraceptive use would be consistent with a bargaining power explanation. That is, when husbands and wives have both been schooled or Ainsworth, Beegle, and Nyamete 115 Table 7. Classification of Countries According to the Marginal Effects of a One-Year Increase in Female and Male Schooling on Current Use of Contraception and Female-Male Schooling Differentials Husband's scbooling Woman's schooling Significant Not significant Significant Kenya [10.21 Botswana [-7.51 Nigeria [9.21 Cameroon [5.1] Senegal [1.01 Ghana [9.71 Tanzania [14.8] Mali [-1.3] Togo [18.31 Niger [-1.61 Zambia [10.1l Uganda 122.1] Zimbabwe [3.9] Not significant Burundi 119.6] Note: Countries are classified on the basis of marginal effects of woman's and husband's schooling from a logit regression on the current use of contraception, using a linear specification of schooling. The samples include only women who have ever been married. Values in brackets are the difference in the percent of ever-married women and men with no schooling. Source: Authors' calculations based on DHS data. only the wife has schooling, then only the wife's schooling makes a differ- ence for contraceptive use; when the husband is more likely to be schooled than the wife, both the husband's and the woman's schooling matters. How- ever, in the latter instances, it is still the wife's schooling that has a larger impact on contraceptive use. This hypothesis and others deserve additional research at the micro level. However, it would not explain as well the differ- ences in the impact of the husband's compared with the woman's education on fertility, presented earlier. Botswana, Ghana, and Zimbabwe all have a very high proportion of female-headed households (45 percent in Botswana, 37 percent in Ghana, and 33 percent in Zimbabwe), which might explain why husband's schooling is not significant (Ayad and others 1994; Ghana Statistical Service 1994). However, this is also true in Kenya (33 percent of households are female-headed), where husband's schooling still affects con- traceptive use (Kenya Central Bureau of Statistics 1994). The results for four countries are graphed in figure 5. The results for the countries not shown generally resemble those for Nigeria. In some countries the relative importance of female compared with male schooling in determining contraceptive use among ever-married women switches according to the level of schooling. The husband's schooling has a larger predicted impact at low levels of male and female schooling, but the wife's schooling has a stronger relation at higher levels of schooling. In Kenya this crossing point seems to be at about three years of schooling, but in Botswana and Zimbabwe it seems to be at about five years and six years, respectively, although husband's schooling is not statis- tically significant in the latter two countries. At higher levels of female school- ing in Kenya and Zimbabwe, the slope of the relationship becomes quite steep for female schooling, but is flat for the husband's. 116 THE WORLD BANK ECO)NOMIC: REVIEW, VOL. 10, NO. I Figure 5. Predicted Relation between Current Use of Contraception and Woman's anzd Husband's Schoolinig, Samnple of Ever-married IWlomen Nigeria Kenya Probability of contraceptive Lise Probability of contraceptive uise .8 - .7 . .6 .0 .5 .5 .4 .4 .33 ..3........ .2 0 2 4 6 8 10 12 1416 0 2 4 6 810 1214 16 Years of schooling Years of schooling Zinibabwe Botswana Probability of contraceptive Lise Probability of contraceptive use .8 . .7 .7- .6 .6 - .5 .5 - 4 A= - D01D= .33- .2 2 .1 .1 0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16 Years of schooling Years of schooling _ Woman - -------- Husband Abte: The graphs show predicted contraceptive Lise for the sample of ever-married women in four countries, presentedl in order of 1991 GNI' per capita, from lowest to highest (see table 2). Results for other cotLntries generally resemble those for Nigeria. a. Husband's schooling is not statistically significant in Botswana or Zimbabwe. Source. Authors' calculations from L)IIS data. VI. CONCLLtSIONS In this article we have analyzed the relations between female schooling and fertility and between female schooling and use of modern contraception in four- teen Sub-Saharan African countries where Demographic and Health Surveys have been conducted. Total fertility rates are high in these countries-ranging from a low of 5.0 in Botswana to as much as 7.4 in Niger. Modern contracep- tive use is below 10 percent in all but Botswana, Kenya, and Zimbabwe. Despite rising female enrollments in many countries, the level of completed schooling among women of reproductive age is extremely low. In six of the countries, half or more of all women have never been to school. Average schooling is highest in Zimbabwe (six years); in four countries the average is less than two years. Ainsuworth, Beegle, and Nyamete 11 7 We find strong support for the negative correlation between female schooling and cumulative fertility in virtually all of the countries, in both urban and rural areas. However, the relationship is nonlinear. The first three years of primary schooling have no relation with fertility in twelve of the fourteen countries. Children are unlikely to become literate in the first three years, and given the state of schooling systems on the continent, it is unlikely that a single year or two of schooling would be enough to radically transform a child's world out- look. Thus, we interpret the two instances of a significant positive effect of early primary schooling on fertility as evidence of some type of selectivity operating among those who dropped out with only a year or two of schooling, rather than as a schooling effect. The last years of primary schooling have a negative rela- tion with fertility in about half of the countries and no relation in the remainder. Secondary schooling is universally associated with lower fertility, even in couIn- tries with less-vigorous family planning programs. Jejeeboy (1992:3), reviewing the evidence on education and fertility prima- rily from the World Fertility Survey, characterized the evidence for "a uniformly inverse relationship" in the poorest countries as "shaky": "a little education appears to lead to higher fertility and we are likely to observe a curvilinear or reversed U-shaped relationship." In fact, she points to an inverted-U shape in ten of twelve African countries, implying that at low levels of schooling fertility rises. In this article, we use different and more recent data sets, control for many more factors simultaneously (including wealth), and use microanalysis of actual fertility (as opposed to the TFR). We find very limited evidence of a positive relation between female schooling and fertility in these countries, and at such low levels of schooling that we question the line of causation. The most intriguing difference among the results for women's schooling and fertility is that for half of the countries upper primary schooling has a negative relation with fertility, and it has no effect in others. Could this reflect differences in the quality of instruction (with a time lag), differences in the labor markets that affect the returns to upper primary schooling, or perhaps differences in the availability of family planning? Unfortunately, exogenous measures of these explanatory variables are not among those available in the individual data sets. Internationally comparable data at the national level are hard to come by and are of varying quality. In additional country-level regressions (with a sample size of fourteen), the infant mortality rate, (,NP per capita, percentage of female- headed households, and various representations of the distribution or average levels of female schooling were not statistically significant predictors of those countries in which four to six years of primary schooling has a negative relation with fertility. Additional research to explain these differences would be very useful to schooling and population policies. Female schooling has also been found to greatly raise the likelihood of con- traceptive use, even among women with primary schooling only. However, as with fertility, the relation is nonlinear. While women with higher levels of school- ing are increasingly more likely to use contraceptives, an important finding is 1 18 IHF WORLL) BANK FCONOMIC: REVIF W. VOIL 10. NO. I that often the marginal relation between an additional year of female schooling and contraceptive use is greatest at the primary schooling level. Again, under- standing why this is observed in some countries and not others, and the relation of these results to the availability of contraception, could lead to new policy insights. Another important difference to be explained across countries is the relative impact of male and female schooling on fertility and contraceptive use, among the sample of ever-married women. Husband's schooling has no significant rela- tion with fertility in about one-third of the countries. In countries where both women's and men's schooling matter, women's schooling exerts a much larger negative effect on fertility than does men's schooling. The analysis also con- firmed that female schooling effects are lower when the samples are restricted to ever-married women. This means that studies that are based on married women understate the effects of schooling on fertility and contraceptive use. Husband's schooling is associated with higher contraceptive use in only half of the fourteen countries. In cases where men's schooling is statistically significant, it generally exerts a smaller influence than female schooling. These results are additional evidence of the importanice of investing in female education to lower fertility and raise contraceptive use. However, attempts to explain why only female schooling matters in some countries but both male and female schooling matter in others were not successful, either for fertility or contraceptive use.11 This study used multivariate analysis of cross-sectional data to examine mar- ginal relationships-such as the relation in the cross-section of altering school- ing by a small amount. However, the levels of female schooling in these coun- tries are so low that more than just marginal increases in female schooling will be necessary. The multivariate regression results do not help us to infer the likely impact of large chaniges in female schooling-such as ensuring that the 40 percent of women with no schooling complete seven years of primary educa- tion. Further, other factors held constant in these regressions-like income- might also change as a result of large increases in schooling. The experience of Botswana, Kenya, and Zimbabwe-where major investments in schooling and family planning have been made in the past decade-may be better indicators of the likely effect of similar policies in other countries. These investments will improve the quality of life for women and children and enhance their future contribution to development, in addition to lowering fertility. 11. Regressors included iNI' per capita, the infant mortality rate, the difference between the percentage of males and females with no education, the percentage of female-headed households, and various measures of the level and distribution of female schooling. Unfort Unately, there are no good measures of exogenous availability of family pllanninig services across countries that do not somehow incorporate actual levels of contraceptive uise. Ainsworth, Beegle, and Nyamete 119 Table A-1. Notes on the Sampling and Coverage of Data Sets from the Demographic and Health Surveys for Fourteen Sub-Saharan African Countries Country Description of oversampling andci weighting Coverage Botswana Oversampling of urbani areas bK a factor of two. National Self-weighted sample withini urban and rural areas. Burundi Oversampling of urban areas by a factor of five. National Self-weighted within urban and rural areas. Cameroon YaoLMd&/Douala oversampled bh a factor of two. National Other urban areas are also oversampled. Self- weighted sample within each uirban and rural stratum. Ghana No oversampling. Self-weighted. National Kenya Oversampled rural areas in fifteen districts. Excludes North Eastern prov- ince and four northern districts accounting for less than 4 per- cent of the national population. Mali Oversampled urban areas. Self-weighted within 100 percent coverage of urban urban and rural strata. areas; 90 to 95 percent coverage of ruiral areas. Nomadic rural population of Timbuktu and Gao excluded. Niger Oversampled Niamey by a factor of four and other All departmients except pastoral urban areas by a factor of three, relative to rural areas of northern desert (zone of areas. Arlit in the Agadez department, arrondissenient of Bilma). Ex- cluded population is less than I percent of national population. Nigeria Oversampled urban areas bv a factor of two. Self- National weighted sample within urbail and rLural strata. Senegal No (iversampling. Self-weighted. National Tanzania Stratified by urban or rural area and by region: National (Mainland and different weights apply to each region and uirban Zanzibar) or rural areas. Togo No oversampling. Self-weighted. National Uganda Urbani areas oversampled by a factor of three. Excludes nine of thirty-four Self-weighted within urban areas. South West districts, with 20 percent of the Region and Luwero Triangle in Central Regioni nationial population. oversampled in rural areas. Zambia Oversamplinig of Luapula, North-Western, and National Western provinces. Zimbabwe No oversamTl pliing. Self-weighted. National Soturce: DHS couintry reports. 120 THE WORID BANK ECONOMI( REVIFW, VOL. 1(0, NO). I REFERENCES The word "processed" describes informally reproduced works that may not be com- monly available through library systems. Ahn, Namkee, and Abusaleh Shariff. 1994. "A Comparative Study of Fertility Determi- nants in Togo and Uganda." International Family Planning Perspectives 20(1):14- 17. Ainsworth, Martha. 1989. Socioeconomnic Determinants of Fertility in Cote d'lvoire. LSMS Working Paper 53. Washington, D.C.: World Bank. - 1990. "The Demand for Children in C6te d'lvoire: Economic Aspects of Fertil- ity and Child Fostering." Ph.D. dissertation. Yale University, Department of Eco- nomics, New Haven, Conn. Processed. Ainsworth, Martha, Kathleen Beegle, and Andrew Nyamete. 1995. The Impact of Fe- male Schooling on Fertility and Contraceptive Use: A Study of Fourteen Sub- Saharan Countries. LSNMS Working Paper 110. Washington, D.C.: World Bank. Alam, lqbal, and J. B. Casterline. 1984. Socio-Economic Differentials in Recent Fertil- ity. WFS Comparative Studies 33. Voorburg, Netherlands: International Statistical Institute. Anker, Richard. 1985. "Problems of Interpretation and Specification in Analyzing Fertil- ity Differentials: Illustrated with Kenyan Survey Data." In Ghazi Farooq and George Simmons, eds., Fertility in Developing Countries. London: Macmillan Press. Anker, Richard, and James C. Knowles. 1982. Fertility Determinants in Developing Countries: A Case Study of Kenya. Liege, Belgium: Ordina Editions. Ayad, Mohamed, Andrea Piani, Bernard Barrere, Koffi Ekouevi, and James Otto. 1994. "Demographic Characteristics of Households." Demographic and Health Surveys Comparative Studies 14. Macro International, Inc., Calverton, Md. Processed. Beegle, Kathleen. 1995. The Quality anld Availability of Family Planning Services and Contraceptive 11se in Tanzania. LSMS Working Paper 114. Washington, D.C.: World Bank. Benefo, Kofi Darkwa, and T. Paul Schultz. 1994. Determninants of Fertilitv and Child Mortality in Cote d'Jvoire and Ghania. LSMS Working Paper 103. Washington, D.C.: World Bank. Bongaarts, John, Odile Franik, and Ron Lesthaeghe. 1984. "The Proximate Determi- nanits of Fertility in Sub-Saharan Africa." Population and Development Review 10(3):51 1-37. Burafuta, Jean-Paul, and David Shapiro. 1992. "Les determinants de la f&ondite au Burundi: Analyse des donnees de l'Enquete Dlmographique et de Sante du Burundi (1987)." Universite de Burundi, Centre Universitaire de Recherche pour le Develop- pement Economique et Social (CURDES). Bujumbura, Burunidi. Processed. Caldwell, John C., and Pat Caldwell. 1987. "The Cultural Context of High Fertility in Sub-Saharan Africa."' Population anid Development Review! 13(3):409-37. - . 1990. "High Fertility in Sub-Saharan Africa." Scientific American 262(May): 118-24. Casterline, John B., Sushella Singh, John Cleland, and Hazel Ashurst. 1984. The Proxi- mate Determinants of Fertility. WFS Comparative Studies 39. Voorburg, Netherlands: International Statistical Institute. Castro Martin, Teresa. 1995. "Women's Education and Fertility: Results from Twenty Demographic and Health Surveys." Studies in Fam7tily Planning 26(4):187-202. Ainsworth, Beegle, and Nyamete 121 Chernichovsky, Dov. 1985. "Socio-Economic Correlates of Fertility Behavior in Rural Botswana." In Dov Chernichovsky, Robert E. B. Lucas, and Eva Mueller, The Household Economy of Rural Botswana. World Bank Staff Working Paper 715. Washington, D.C. Cleland, John, and German Rodriguez. 1988. "The Effect of Parental Education on Marital Fertility in Developing Countries." Population Studies 42:419-42. Cochrane, Susan Hill. 1979. Fertility and Education: What Do We Really Know? World Bank Staff Occasional Papers 26. Baltimore: Johns Hopkins University Press. . 1988. "The Effects of Education, Health and Social Security on Fertility in Developing Countries." wrs 93. World Bank, Washington, D.C. Processed. Cochrane, Susan H., and Samir Farid. 1990. "Socioeconomic Differentials in Fertility and Their Explanation." In George T. F. Acsadi, Gwendolyn Johnson-Acsadi, and Rodolfo Bulatao, eds., Population Growth and Reproduction in Sub-Saharan Af- rica: Technical Analyses of Fertility and Its Consequences. A World Bank Sympo- sium. Washington, D.C. Fairlanib, Cheryl D., and Wilhelmus L. Nieuwoudt. 1991. "Economic Factors Affecting Human Fertility in the Developing Areas of Southern Africa." Agricultural Econom- ics 6:185-200. Farooq, Ghazi M. 1985. "Household Fertility Decision-Making in Nigeria." In Ghazi Farooq and George Simmons, eds., Fertility in Developing Countries. London: Macmillan Press. Fenn, Thomas, Therese McGinn, and Yves Charbit. 1987. "Direct and Indirect Det- erminants of Fertility in Burkina Faso." Paper presented at the Annual Meetings of the Population Association of America, Chicago, April 30-May 2. Processed. Feyisetan, Bamikale J., and Martha Ainsworth. 1994. Contraceptive Use and the Qual- itv, Price, and Availability of Family Planning in Nigeria. LSMS Working Paper 108. Washington, D.C.: World Bank. Ghana Statistical Service and Macro International, Inc. 1994. "Ghana Demographic and Health Survey 1993." Accra, Ghana, and Calverton, Md. Processed. Greene, William H. 1981. "On the Asymptotic Bias of the Ordinary Least Squares Estimator of the Tobit Model." Econometrica 49(2):505-13. . 1993. Econometric Analysis. New York: Macmillan Press. Jejeeboy, Shireen. 1992. "Women's Education, Fertility and the Proximate Determi- nants of Fertility." Paper for the Expert Group Meeting on Population and Women. Gaborone, Botswana. June 22-26, 1992. International Conference on Population and Development 1994 (UNFPA). Document ESD/P/ICPD.1994/EG.Jn/13. Processed. Kelley, Allen C., and Charles E. Nobbe. 1990. Kenya at the Demographic Turning Point? World Bank Discussion Paper 107. Washington, D.C. Kenya Central Bureau of Statistics, National Council for Population and Development, and Macro International, Inc. 1994. "Demographic and Health Survey." Nairobi, Kenya, and Macro International, Inc., Calverton, Md. Processed. Lockheed, Marlaine E., and Adriaan M. Verspoor. 1991. Improving Primary Educa- tion in Developintg Countries. New York: Oxford University Press. Maddala, G. S. 1983. Limited-Dependent and Qualitative Variables in Econometrics. New York: Cambridge University Press. Montgomery, Mark, Aka Kouame, and Raylynn Oliver. 1995. The Tradeoff between Numbers of Children and Child Schooling: Evidence from C6te d'lvoire and Ghana. LSMS Working Paper 112. Washington, D.C.: World Bank. 122 THE WORLD BANK FCONOIMIC REVIEW, VOl 0), NO. I National Research Council, Committee on Population. 1993. Factors Affecting Con- traceptive Use in Sub-Saharan Africa. Washington, D.C.: National Academy Press. NjoguL, Wamucii. 1991. "Trends and Determinants of Contraceptive Use in Kenya." Demography 28( 1):83-99. Okojie, Christiana E. E. 1990. "Women's Status and Fertility in Bendel State of Ni- geria." Economic Growth Center Discussion Paper 597. Yale University, New Ha- ven, Conn. Processed. 1991. "Fertility Response to Child Survival in Nigeria: An Analysis of Microdata from Bendel State." In T. Paul Schultz, ed., Research in Population Economics. Vol. 7. Greenwich, Conn.: JAI Press. Oliver, Raylynn. 1995. Contraceptive Use in Ghana: The Role of Service Availability, Quiality, and Price. i.ssiss Working Paper 111. Washington, D.C.: World Bank. Pitt, Mark. 1995. Women's Education, the Selectivity of Fertility, and Child Mortality in Sub-Saharan Africa. LSMS Working Paper 119. Washington, D.C.: World Bank. Schultz, T. Paul. 1992. "Assessing Family Planning Cost-Effectiveness: Applicability of Individual Demand-Program Supply Framework." In J. R. Phillips and J. A. Ross, eds., Family Plinning Programmes and Fertility. Oxford, U.K.: Oxford University Press. . 1994. "Sources of Fertility Decline in Modern Economic Growth." Yale Uni- versity, Department of Economics, New Haven, Conn. Processed. Scribner, Susan. 1995. Policies Affecting Fertility and Contraceptive Use: An Assess- ment of Twelve Sub-Saharan Countries. World Bank Discussion Paper 259. Wash- ington, D.C. Shapiro, David, and Oledo Tambashe. 1994. "Education, Employment, and Fertility in Kinshasa and Prospects for Changes in Reproductive Behavior." Penn State Depart- ment of Economics Working Paper. Pennsylvania State University, State College, Pa. Processed. Snyder, Donald W. 1974. "Economic Determinants of Family Size in West Africa." Demography 11 (4):613-27. Thomas, Duncan, and John Maluccio. 1995. Contraceptive Choice, Fertility, and Pub- lic Policy in Zimbabwe. i,sM%is Working Paper 109. Washington, D.C.: World Bank. United Nations. 1987. Fertility Behavior in the Context ofDevelopment: Evidence from the World Fertility Survey. Population Studies 100, ST/ESA/SER.A//100. New York. van de Walle. Etienne, and Andrew Foster. 1990. Fertility Decline in Africa: Assess- mients and Prospects. World Bank Technical Paper 125. Washington, D.C.: World Bank. World Bank. 1984. World Development Report 1984. New York: Oxford University Press. 1986. Population Grouwth and Policies in Sub-Saharan Africa. Washington, D.C. 1992. World Development Report 1992: Development and the Environment. New York: Oxford University Press. . 1993. World Development Report 1993: Investing in Health. New York: Ox- ford University Press. . 1995. World Development Report 1995: Workers in an Integrating World. New York: Oxford Universitv Press. Fertility and Child Mortality in Cote d'Ivoire and Ghana Kofi Benefo and T. Paul Schultz This article examnines individual, household, and commnunity characteristics that may affect fertility in contemporary 05te d'livoire anid Ghbana and the relationship be- tuween child mortality and fertility. It was not possible to reject the null vypothesis that child mortality is exogenouss. Treating child niortality ais exogenous, fertility responds directly to c-hild mortality. but by ai snmaller proportioni thba estimated in studies of East Asia andt lIatin America. Increases in feniale education and urbanizal- tion are likely to contribute to declines in fertility itn both cointries, hut economic grow-th uwithout these structural changes is not yet stronigly, related to lower fertility. This article examines individual, household, and community characteristics that affect fertility in two neighboring West Africani countries: C6te d'lvoire and Ghana. It analyzes the relationship betweenl child mortality and fertility and examines the idea that high levels of child mortality encourage parents to have large numbers of births (Notestein 1945; Smith 1961; Freedman 1975; Schultz 1969, 1976). The reduction in child mortality is an obvious objective of parents and society. Public programs that promote child health mighit, nonetheless, re- ceive still more support if reduced levels of childL mortality were shown to coI1- tribute to reducing fertility and thereby to slowing population growth. There are conceptual and statistical problems with measuring the causal rela- tionship between child mortality and fertility. Both variables may affect each other, both may be modified by other factors, and both may be measured with error, generating a difficult-to-interpret association between fertilitv and child mortality. For social scientists to measure without bias the effect of child mor- tality on fertility, they must observe features of the woman's environment that affect only the mortality of her own child. One such environmental factor might be a local public health program that increases child survival, but does not oth- erwise directly affect fertility. Kofi Benefo is with the Department of Sociology at Browvn nliiiversityv and T. Paul Schultz is withl the Department of Economics at Yale tiniversity. This article was writtenl as hackgroUnd tor the research project on "The Economic and Policy Determilants of Fertility in Suh-Saharan Africa," financed by the World Bank Research Committee (RPO 67691) and sponsored bv the Africa Techinical Department and the Policy Research Department of the World Banhk. he authors acknowledge helptul comments by the participants at a World Bank workshop: commenits hy Ann loidd. Mark Montgomery, Ravvivl Oliver, azid the anilVymoLIs referees; and the encoLuragemiienit and cominmelnts hv Martha Ainsworth. (C 1996 The International Bank for Reconstruictioni and Development /TiII VORiLD BANK 12 11 124 1H F. WORLD BANK ECON OMIC REVIEW, VO'I. 10, NO. I Most previous studies of fertility in Africa have relied on surveys such as the World Fertility Survey or the subsequent Demographic Health Surveys, neither of which collected much information on the household's economic characteris- tics, community environmental setting, or local health programs (Barbieri 1989). This article analyzes data from the Living Standards Measurement Surveys (LSIMS) in C6te d'lvoire and Ghana, which were collected in the late 1980s, and include information on household consumption and economic behavior, as well as on prices and conditions in each sample cluster (or community). These data have greater potential for clarifying the economic and educational determinants of fertility in combination with child mortality than do those from previous Afri- can surveys (Ainsworth 1989). Section I discusses the determinants of child mortality and fertility and ex- plores the statistical problems that arise in estimating interrelationships between the two. Section II presents the methodological issues, including data sources, sample selection, variable definitions, and descriptive statistics. Section III re- ports estimates of the determinants of child mortality and fertility. Section IV presents conclusions. An expanded version of this article appears in Benefo and Schultz (1994), and a contemporary overview of demographic conditions in Sub- Saharan Africa is provided by van de Walle and Foster (1990). 1. DFTERMINANTS OF FERTILITY AND CHILD MORIALITY This section develops the conceptual framework for the choice and treatment of determinants of fertility and child mortality. Exogenous determinants of fertil- ity are characteristics of the woman, her household, and community over which she exerts no control, but which are likely to affect her fertility. By contrast, endogenous variables that may determine a woman's fertility reflect in part deci- sions and choices that either she or the members of her household have made and that are constrained by her economic endowments and luck. This distinc- tion between exogenous and endogenous variables is important, because the use of endogenous choice variables to explain fertility and child mortality can lead to biased and misleading results if this endogeneity is not recognized and prop- erly treated econometrically. This conceptual framework could lead to two alternative but potentially com- plementary approaches in an empirical analysis of fertility. The first approach estimates reduced-forn relationsbips, including among the explanatory vari- ables all exogenous variables that potentially affect fertility, directly or indi- rectly. This approach thereby excludes family outcome variables that are endogenous because they may be affected by the woman's behavior or choices. In other words, because the errors embodied in such endogenous variables are likely to be correlated with those in the fertility equation, they must be excluded from reduced-form relationships to avoid bias by single-stage estimation meth- ods, such as ordinary least squares (oLs). The correlation of errors could be caused by preferences, unobserved heterogeneity of couples, or other omitted Benef)t and Scbhultz 12 ) factors and measurement errors. Parallel reduced-forimi analyses of child mor- tality determinants can he undertaken. The second estimation strategy is pursued here and introduces further struc- tural assumptions about how some endogenous family outcome variables, such as the child death rate, are themselves determined by additional variables that do not directly impact fertility. Thiese structural assumptions are used to iden- tify statistically and to estimate consistently the effect of an endogenous ex- planatory variable on fertility, conditional on the hypothesized structure. Fertility, lIzcome, andi Cbild Mortaliti- Economic theories of fertility assume that parents have the number of chil- dren they want, given the costs and benefits associated with having a specific number of births, including the costs of avoiding unwanted births. The lifetime demand for births (F) is a function of many socioeconomic factors, and we stress a few of the lifetime economic constraints as thev affect a representative woman: the woman's productive opportunities, W; her household's nonhumani capital assets, V ; the mortality rate her chilldreni experience, D ; local market prices, P ; and local public services and environimienit, .S (Willis 1974; Ben Porath 1978; Schultz 1981); (1) F = F [W(E, T), V, D; P, SI where E denotes the woman's educationi and T her height. The productivity of a woman's time is expected to decrease her demand for births, because the increase in the opportullity cost of her time in childcare out- weighs, on balance, the increase in her incomiie opportunities (Schultz 1981). The woman's schooling may have many effects on her opportunllties and behav- ior, one of which is to increase her wage. Edducationi may also affect women's demand for childreni in other unspecified ways and shift their biological supply of births, positively through improvements in their health and nutrition, and perhaps also through reducing their risk of contracting sexually transmitted diseases, which are implicated in premature sterility and a shiortfall in reproduc- tive supply. The educational control variables are thus proxies for the produc- tive value of a woman's time, her ability to practice effective birth colitrol, and all other mechainisms by which education influences reproductive capacity, goals, and behavior (Kritz and Gurak 1989). In both C(,te d'lvoire and Ghana, the education of women is strongly associated withi their wages. The proportionate gain in wages is larger at the secondary school level than at the primary level per year completed, and wage returns to schooling are generally larger in Cote d'lvoire than in Ghana (Schultz and Tansel 1992. Consequently. education is specified as a spline in years completed at three levels of schooling. Taller individuals are widely observed to be more productive and to have fewer health problems. The nutritional status that is achieved throughi improved diet and reduced exposure to disease durinig early childhood increases the heigilt that an adult can reach (Fogel 1991; Schultz 1995). Migrationi from a rural 126 THE WORLD BANK ECONOMIC REVIEW, VOL. 1t, NO. I region to an urban area is another important investment that increases the produc- tivity of women (by 21 to 64 percent in Ghana and C6te d'Ivoire, respectively) (Schultz 1995: table 2). However, adult migration is more likely to be determined simultaneously with fertility. Consequently, migration is endogenous in a life- time fertility model, and it is omitted from this analysis. Results in which migra- tion is treated as exogenous are reported elsewhere (Benefo and Schultz 1994). Market income per adult in the household is endogenous to lifetime cumula- tive fertility because it embodies the effect of labor supply and family composi- tion decisions, and it is also intertemporally redistributed by savings and invest- ments, such as the choice as an adult to migrate from a rural to an urban area. Most clearly, the woman's market labor supply and her fertility over the life cycle are jointly determined, to the extent that children and market labor place competing demands on her budget of time. What initial economic conditions are expected to affect a woman's lifetime demand for children? The framework calls for information on her inherited nonhuman capital assets or exogenous cumulative transfers, just as we have included indicators of her initial human capital stocks as an adult, in the form of education, and early nutrition, proxied by height. Without information in our data on inherited nonhuman capital assets or exogenous transfers, we can either omit altogether income from the model or use current assets of the household per adult as an instrument to predict household income (proxied commonly by more accurate measures of consumption per adult). This latter estimation approach could be misleading if current assets are themselves endogenous to the life-cycle accumulation process, in which case assets would be an invalid instrument. Elsewhere we have reported these instru- mental variable (IV) estimates that treat household income or consumption as endogenous (Benefo and Schultz 1994). In this article, at the urging of a referee, we have eliminated income or consumption entirely from the model, and im- plicitly income is solved out of the partially reduced-form equation estimated. The household assets per adult and the prevalence of tree crops in the com- munity are retained from the previous study as exogenous conditioning vari- ables. The notable finding of that previous study was that income was positively associated with fertility in C6te d'lvoire and negatively in Ghana, and positively in rural areas of both countries and negatively in urban areas. The relative prices of market goods and public services in the community may have diverse effects on the costs and benefits of fertility, and are in- cluded as control variables. If parents want to have fewer children than they would otherwise have, the price of contraception should affect fertility. Un- fortunately, insufficient information on the local availability and cost of birth control methods is available to include them in this study (Benefo and Schultz 1994). Family planning services are not provided by the public health system in C6te d'lvoire, and the LSMS data for that country do not include individual- or community-level information on birth control practices or availability of family planning methods. A health facility questionnaire collected informa- Benefo and Schultz 127 tion in Ghana about the provision of family planning services and availabil- ity of supplies, but these facility data could be matched with only 60 percent of the LSMS households. About 2 percent of women report using modern methods of contraception in Cote d'lvoire and 5.6 percent in Ghana (Oliver 1995). The variation in fertility analyzed in this study appears to be largely associated with differentials in the timing and incidence of marriage and traditional birth-spacing practices. The first model specification assumes that all of the explanatory variables in the fertility equation are exogenous. OLS estimates of model I are a reasonable first step in the empirical analysis of the survey data on these variables.' The decision of a woman to marry-and consequently the characteristics of her husband-or to head her own household can be viewed as decisions made jointly with fertility over the life cycle. If these household composition variables are themselves adapting to changes in the constraints on individual choice dur- ing the development process, it is not clear how they would be structurally iden- tified in a fertility model (Schultz 1994). A common practice of stratifying the population and explaining the reproductive behavior of only married women merely transforms the problem into one of correcting for sample selection bias, which also depends on some form of identification. One solution to this conun- drum is to omit the household composition and husband characteristics from the model entirely and implicitly solve them out of partially reduced-form equa- tions as with income. This approach would seek to understand the woman's behavior on the basis of her own initial endowments and locally evolving oppor- tunities. This specification is estimated as model 11. The specification for model III makes the less satisfactorv assumption that household composition and husband characteristics are exogenous to fertility and are likely to affect child mortality and fertility. Most economic studies of fertility are based on a static formulation of a life- time demand for births, which abstracts from considerations of dynamic opti- mization. Without a panel survey of long duration that measures time-varying constraints on fertility, our model can only attempt to account for lifetime fertility, conditional on nonlinear age effects. Child Mortalitv Child mortality can affect a woman's demand for births in two ways. First, it can induce her to replace ex post her children who die. This response could occur by means of a biological feedback or through adaptations of behavior. This mechanism would be more effective if childbearing is initiated at an early age and premature sterility is infrequent, allowing most couples to have the biologi- cal capacity to bear more births than they want. Even if parents do not use mod- ern birth control practices, shorter periods of postpartum abstinence and shorter 1. Tobit and ordered probit models can also be estimated to deal with the censored and discrete form of the fertility variable. Reestimation of our final models did not noticeably change the conclusions reviewed here. 128 lilt WD'ORID R\NK -( O)N('NII( RlVIF\X1 \ (1. 10., N( I durations of breastfeeding could allow the fertility of individuals to compensate substantially for their experience of child mortality. Second, in a society where child mortality has been stable or slowly declining for some decades, parents can adapt their fertility behavior in anticipation of the levels of child mortality they will experience on1 average. One form of antici- patory behavior would be the case in which parents help their children marry earlier in higher child mortality regions, as observed in Taiwan (China), for example (Schultz 1980). Early marriage might be preferred to shortened birth intervals in environimients where early weaninig of infants would expose them to further health risks. An objective of this article is to assess the sum of the replacement and antici- patory responses of fertility to child mortality. As indicated in the introduction, estimating this effect from the observed association may be misleading. First, the causal effects canl flow in both directions, overstating the one-way effect we want to estimate. High levels of fertility may increase child mortality by stretch- ing family resources and the biological capacity of the woman to bear healthy children. Second. unobserved features of the woman, her household, and her community may contribute to her experiencing higher child mortality and fertil- ity, introducing a spurious correlation between the two outcomes. Third, errors in measuring the appropriate (realized and anticipated) child mortality would probably bias downward the estimated effect of child mortality on fertility. Fourth, the possible values of the child mortality rate depend on the number of children born, a factor that introduces additional forms of interdependence (Olsen 1980). Child mortality may be affected by all the exogenous variables in the fertility equationi (equation 1) and by certain additional exogenous factors that change the relative cost or availability of child health inputs for the parents. We assume that a set of community variables (C), associated with water and sanitation infrastructure, distance to the nearest health clinic, and community disease prob- lems (Patterson 198 1; Morrow, Sniith, and Nimo 1982) affect the proportion of children dying before their fifth birthday, D, hut do not otherwise impact a womani's fertilit%: (2) D = DIW(E, T), V; P, .S, Cl. Household assets are expected to reduce child mortality. Most recent studies of child mortality find a woman's education to be related negatively to her ex- perience of child mortality, although this effect of female education could be due partly to the child health inputs she is able to purchase or produce with her education-enlhaniced wages and improved marriage prospects (Pitt 1995). The mother's height is a proxy for her own health investmiients as well as her genetic health endowments, whichi are likely to improve the chances for her children to survive. With child mortality, errors in measuremenit may be substantial. By treating child mortality in the fertility equation as endogenious and estimating it by in- Benefo and Schultz 129 strumental variables (IV), we should eliminate both simultaneous equations bias and classical random measurement error bias. Hausman (1978) specification tests can also be implemented to test whether child mortality is empirically en- dogenous, as we hypothesize in models 11 and III. The IV estimates of fertility (equation 1) and the Hausman specification tests will be treated with greater confidence if the identifying instruments in the child mortality equation (equa- tion 2) are jointly significant in explaining child mortality, after controlling for the other exogenous variables. Additional Issues of Empirical Specification The unusually high level of child mortality between the ages of one and five in West Africa is an argument for basing our analysis on a five-year cohort mea- sure of child mortality rather than a conventional infant mortality rate. Because weaning is often delayed until after the first year, it seemed important to under- stand how household resources and environment affect child survival prospects beyond the critical weaning-feeding transition, when children must develop their own immunities to local pathogens and local water and sanitation facilities may affect child mortality. In order to measure this dimension of child mortality, our working sample must be limited to women who have had a child five years before the survey. The choice of five-year child survival rates (q in lifetable terms) is perhaps more suitable than infant mortality if the goal is to estimate replacement re- sponses to child mortality. The analysis was repeated using only infant (first year) mortality as our measure of child deaths for the same sample examined here, and again for the larger sample that included fertility and infant mortality for all births up to one year before the survey. No substantial changes in the conclusions were noted, although food prices tended to be more significant in explaining child mortality through age five than they were in explaining infant mortality. The opportunity cost of a woman's time is expected to be an important det- erminant of the price of having children. Although women in West Africa are the main providers of childcare and are often responsible for producing or purchas- ing the food and medical care children receive, women infrequently work as wage laborers. In the late 1 980s, in Cote d'lvoire only 4 percent of women worked for wages, and in Ghana the proportion was 7 percent (Schultz and Tansel 1992). Consequently, any prediction we might devise for the market wage that a par- ticular woman could expect to receive might not be a reliable indicator of the opportunity cost of her time if she worked outside of the wage labor force. In both countries, however, the education of women is strongly associated with their market wage rates and self-employment earnings. Thus, it seems reason- able to use years of educational attainment to proxy the value of women's time in both wage and nonwage work (van der Gaag and Vijverberg 1987). Because the community price series are highly intercorrelated, any one price should not be treated as varying independently of the other prices for the pur- 1,30 THF WORILD BANK EC:ONOMI( REVIFW. V O. 11) NO. I poses of policy simulations. Rather the entire set of prices summarizes the fac- tors responsible for the relative scarcity of basic foods and household staples, such as local climate and geography, transportation infrastructure, and market integration. Finally, characteristics of the local economy and health and sanita- tion system are included to assess the impact of these policies. 11. METHODOLOGIC:AL ISSUES IN THE EmPIRICAL, SPECIFICATION The data for this study are primarily from the L.SS conducted by national statis- tical agencies in collaboration with the World Bank. We use the first three rounds of the Cote d'lvoire Living Standards Survey (CILSS) conducted in 1985, 1986, and 1987, and two rounds of the Ghana Living Standards Survey (GLSS) in 1987-88 and 1988-89 (Ainsworth and Munoz 1986; Glewwe 1987). The CILSS interviewed 1,600 households per year for a total of 4,800. and the G;LSS interviewed 3,200 households per year for a total of 6,400. The surveys used two-stage, self-weighted stratified (by three agroecological zones and size of localities) sample designs. The surveys collected reproductive histories of one randomly selected woman of childbearing age in each household. The working sample is restricted to women who had at least one birth five or more years before the date of the survey. A small number of womeni (or their husbands) who did not report information are also excluded. Our workinig samples include 1,943 women in C6te d'lvoire and 2,237 in Ghana. Each sampling cluster (or enumeration area) contains 16 house- holds. Each year 100 clusters were sampled in C6te d'lvoire and 200 in Ghana. The samples are a rotating panel, with half of the clusters randomly replaced each year. A subset of the sample households is reinterviewed in adjacent years, but these households have not been matched for this study. The household and individual identificationl codes from the first year were not preserved in the second- year data in Ghana; thus matching would be difficult. About 600 households were reinterviewed in Cote d'lvoire in each following year and can be matched. This information is not used to adjust the standard errors in this study because, although the households can be matched, the woman interviewed in each house- hold was not necessarily the same (Schultz 1995). The coefficient estimates ob- tained here are probably not biased by our failure to incorporate this sampling design in our estimates, but the true standard errors may be somewhat larger than those reported. Definitions of Variables The woman is the unit of analysis. The human capital endowment of the woman is summarized by her age, education (years by three levels), and height. If she is married, similar information is available for her spouse. Household- level information includes the sumn of the value of business assets, land, and nonearned income.' Appendix table A-1 reports the means and standard devia- tions of all of the variables for the working samples of womeni. 2. Nonearnied inicome is capitalized witih other assets at i 1) percent rate of return. Benelo and Scbhultz 13 1 Major differences in the timing and extent of social development of the two countries are reflected in their educational systems. In Cote d'lvoire, the colonial and postcolonial governments kept close control over schools. This resulted in a schooling system modeled quite closely after that in metropoli- tan France, with access restricted by com petitive exams to only a small frac- tion of the populatioln. The colonial governmienit in Ghana adopted a largely laissez-faire attitude to education, allowing private school systems to respond to local demands not satisfied by public sclhools. Private and missioniary orga- nizations developed a quite differentiated and geographically dispersed sys- tem of schools. Currently, levels of schooling are higher and gender inequalities in education are smaller in Ghana than in Cote d'lvoire. Thirty-four percent of primary school enrollment in C6te d'lvoire in 1965 was female, compared with 41 percent in Ghana (World Bank 1986). The differences in enlrollmenits by gender widen at the secondary level. Figure 1 presents the distrillutiolnl of schiooling for womliell fifteen years and older for both countries from01 our surveys. In Cote d'lvoire, 72 percent of women have no schooling-54 percenit of those in urban areas and 88 percent of rural women. In Ghana, 42 percenit ot womileni have no sclhooling-33 percent of the womeii in iurban areas and .57 percenit of rural women. In Ghana, 25 percent of all wonien-33 percenit of those in urban areas and about 15 percent of those in rural areas-have ten or more years of schooling. In contrast, in C6ote d'lvoire less than 5 percent of all womlienl have ten or more years of schooling. Community-level (or sample cluster) information is available for rural areas, but often in the case of urban areas the conditions maiv be inferred from othier data sources. Informationi abotit prices of six food stalples in the local market, the distance to weeklv markets, and the proportion of the cultivated land planted in tree crops is averaged for each cluster from the sample responldenlts. AnIual average rainfall for the sample clusters is obtained from the closest weather station.' Rainfall is associated both withi more favorable agricultural produc- tion opportunities and also with the presenice of cer-taini parasitic and infectious diseases, such as malaria, that are responsible for a substantial share of the deaths of children ltider age five in West Africa (Feachem and jamison 1 991). Other specific comMunity-level variables that are assumred to directly affect child health include the distance to a health cliniic; the percentage of households with running or protected water supplies and sanitaltioni facilities (toilets or latrinies); and malaria, diarrhea, or measles being one of the two most serious health prob- lems in the community, accordinig to the comruinu.1ity respondenit. Additional inforimiation is also obtained from the (, ss on whether there had been a child immunization campaign in the localitxv in the five years before the survey. The governmient budget for Ghana in 1988XX WS used to estimate public expenditures per capita oin healthi programs (exclusive of doctor trainling) in each of the ten 3. Providedi 1y .a research a s,istalnt oft A igois D)eaton ti ir ( OrUC L'Iv),. 1 ad deri vd r i -oil vi i weather mlaps for Ghana. 132 THE WORLD BANK ECONOMIC REVIEW, VOL. 10. NO. I Figure 1. Distribution of Women 15 Years and Older by Schooling, CMte d'Ivoire, 1985-87, and Ghana, 1987-89 CMte d'lvoire Ghana Percentage Percentage of total All women of total All women 30 _ 72% 30 25 25 20 20 15 15 10 10 5 5 0 0 0 2 4 6 8 10 12 14 16+ 0 2 4 6 8 10 12 14 16+ Years of schooling Years of schooling Percentage Percentage of total Urban of total Urban 0 2 4 6 8 10 12 14 16+ 0 2 4 6 8 10 1214 16+ Years of schooling Years of schooling Percentage Percentage of total Rural of total Rural 30;'-50 30. 39/ 25 -< 25 . 20 -20 15 - 15 - 10> 10 r 5 ~51 0 2 4 6 8 10 12 14 16+ 0 2 4 6 8 10 12 14 16+ Years of schooling Years of schooling Source Authors' tabulations of survey data. Benefo and Schultz 133 regions of Ghana. These public health expenditure levels are attributed to all clusters in each region in 1987-88 or 1988-89. The woman's regional residence and ethnicity or language and, in the case of Ghana, religion are considered in the analysis as potential determinants of fertil- ity, child mortality, and income. Controlling for these regional characteristics, as well as for climate, agricultural cropping patterns, and endemic disease prob- lems is expected to diminish the estimated effects of individual education and health and community program variables on mortality and fertility. Assets are first deflated to adjust for regional differences in price levels in both countries, and in particular for the higher cost of living in Abidjan and a slightly higher price level in Accra. Then, because the surveys were collected over two to three years, the expenditures reported by the respondent are further adjusted for the national price level during the month of the survey. This real value of the asset is based on the prices prevailing in the first month of the Ghana survey, that is, September 1987, and on the average prices for all of 1985 in Cote d'Ivoire. A monthly price index was not available for C6te d'lvoire, but since the rate of inflation was less than 10 percent per year from 1984 to 1988, it was simply assumed that the annual rate of inflation was uniformly distributed over the twelve months from July of one year to June of the next. Prices are expressed relative to those in the base year. Community Variable Correlations The community data have three problems. First, there are at most only 200 observations on the several dozen community variables. Second, these community variables are highly intercorrelated. For example, the correlation between the distance to doctors and distance to clinics in the working sample of Ghana and Cote d'lvoire is 0.70, and only the latter is included in the model to obtain more stable and precise estimates. The researcher must distill the community vari- ables down into only a few reasonably distinct features of the communities. Consequently, what appears to be a rich array of community characteristics is realistically a much more sparse set of community information. This problem of intercorrelation among community variables limits the capacity of the researcher to assess the effects of individual program interventions. Third, interregional variation in programs and policies may not be indepen- dent of household resources or individual preferences. Health programs may be offered in a region that has a particular health problem. Malaria eradication or child immunization campaigns tend to be fielded in poorer, more remote regions of Ghana, where women are relatively less educated. It should not be surprising, therefore, that some of these types of compensatory public health measures are associated with higher regional levels of child mortality (Rosenzweig and Schultz 1982; Rosenzweig and Wolpin 1986; Schultz and Tansel 1992). In addition to this difficulty of program evaluation when programs are tar- geted to communities with special health problems, migrants within a popula- tion may move on their own accord toward healthier environments and toward 134 rHE U'ORLD BANK ECONOMIC REVIEW, VOL. 10, NO. I regions served by better public health programs. If such migrants also are in- clined to invest more in their own health and that of their children, for reasons that researchers cannot observe, this form of selective migration may also bias conventional regional-based policy evaluation studies using cross-sectional data. To deal with both of these evaluation problems, development agencies might phase their health and welfare pilot programs and policy interventions indepen- dently of other confounding background factors. Then the correlations between the implementation of a new program and household behavior and outcomes can be interpreted more confidently as evidence on the payoff to public policies and expenditures. Descriptive Statistics Our working sample is not representative of the entire population, because women who had not yet had a birth five years before the survey are excluded. Mean fertility in the sample exceeds that in the population, particularly among younger women. Appendix tables A-1 and A-2 in Benefo and Schultz (1994) report the number of children born for all women by age and education for the entire sample to facilitate comparisons with other survey estimates. Figure 2 shows the average number of births per woman for both the working and the entire samples by maternal age and education. The figure shows how the differences between the samples decline as the age of the woman increases. The child mortality rate is a cohort five-year rate; there is little reason to think this rate differs substantially between the restricted and unrestricted samples. Table 1 reports the number of children born and the child mortality rate for the entire working sample of women by schooling and region, and for women age thirty-five to forty-nine. Our discussion covers only the older age group, whose representativeness should not be affected by the restriction that the woman had a birth five years before the survey. Child mortality is about equal in the two countries-16 percent in Ghana and 17 percent in C6te d'lvoire-but fertil- ity is slightly lower in Ghana, 6.13, compared with 6.53 for C6te d'lvoire. In this age group, the effect of one to four years of schooling is associated with an 8 percent decline in fertility in Ghana, compared with women with no educa- tion, and a one-sixth decline in child mortality. In C6te d'lvoire, one to four years of schooling is associated with fertility being 13 percent lower and child mortality being a third lower. Further education tends to be associated with additional declines in both fertility and child mortality, although the limited number of better-educated women, particularly in Cote d'lvoire and in the rural areas, makes the estimates imprecise. Education is a major correlate of lower fertility and child mortality in both countries. One difference between the countries is the age differentials in child mortality, summarized more fully elsewhere (Benefo and Schultz 1994). In Cote d'lvoire, young women with no education appear to have experienced lower child mortality than older women, in both urban and rural areas. In Ghana young women's experience of child mortality is not appreciably different from Benefo aLnd Schultz 1.35 Figure 2. Average Number of Births per Woman, bv laternal Education and Age, C6te d'Ivoire. 198587, antd Ghana, 1987-89 C6te d'Ivoire Gbhanla Average bihrts AvcraMg Ii Irhs per woman All wornen p1 1 v.rIflnin All women 10 I 8 8 6 6 4 2 2 *- 14 l-4-11r1 Ii Al 5 I-) I , I, 5 I',ll 0I Ir' 5 -,r I-) I I I,c rIil+ ,I-I , 1 I. I 1r+ ;0 - 5 nl1-1S0 1 - I ( i 510 1i+ I 2- 2i -Ai AS -10 S 15 21 25 0 Al5./,( Yeals 0,1 s(ch-l Iing mad age grtill Yellr ol sr l-rinllg .1nTd ige gloup Average birilis Aseiagc hulris pel woilarl Urban per -snr,rrrr Urban 1. Ill 88 6 1. ON M I14 S.l() I 1t (- I - s it1 11 (I i 5 1c] I I 1+ s W ; Ic 118 I i c W\ I I1 ( > I , s 5 1 1 .1 I , s 5 ' 1 1 1) I I}s 5 III I 1 + 15 2 0 AI, I i'! l, 1; 2 2-A 55- 10 s- YearMS oII s.r1 1r-oAg 11ad age gouLIp Yc.-r cI sr 11-.d*lin g aolli ig gloupl Average dlirs '%lOs Lcr:ig 1rI111hr pci woman Rur3L p,. Nw-Illrl Rural 8 ~~~~~~~~~~~~~~~8 2 ~~~~~~~~~~ -.~~~~~~~~ W W I 50 II + sl<^c - 1 ] I+ . I s; I| I i+ sl| 11) I I s IC, 1 ll 0 + > 1 I I [1 1 1 i 31 s 1l1. II- I s 1 9 III II- I' 2-1 20 A- As -ir -As A5 1r) 9sri yeais 0 I rch-linig andi age glrrlp Yr.1s rl1 scI( r-rrrng aild Ige grIoup c Wot king sirmpit U RFll sample Sormre Arntirorscalr nnllrrnrr 136 THU. WORLD BANK ECONOMIC REVIEW, VOL. 10, NO. I Table 1. Fertility and Child Mortality, by Women's Schooling and Region, Cote d'Ivoire, 1985-8 7, and Ghana, 1987-89 Women's schooling (years) Country and indicator None 1-4 5-10 1 1 or more All women Ghana Fertility rate' All women Total 5.43 4.79 3.97 3.27 4.73 Urban 5.49 4.71 3.86 3.19 4.54 Rural 5.37 4.93 4.25 4.00 5.01 Women age 35-49 Total 6.67 6.11 5.38 3.70 6.13 Urban 6.59 5.77 5.09 3.58 5.80 Rural 6.75 6.81 6.41 4.60 6.67 Child mortality rateb All women Total 0.19 0.15 0.12 0.12 0.16 Urban 0.16 0.17 0.12 0.12 0.14 Rural 0.22 0.13 0.12 0.09 0.18 Women age 35-49 Total 0.18 0.15 0.12 0.07 0.16 Urban 0.16 0.16 0.12 0.07 0.14 Rural 0.21 0.12 0.13 0.05 0.19 C6te d'lz!oire Fertility rate' All women Total 5.82 4.43 4.13 2.67 5.49 Urban 5.33 4.19 4.03 2.68 4.83 Rural 6.05 4.72 4.55 2.50 5.94 Women age 35-49 Total 6.64 5.79 5.52 3.57 6.53 Urban 6.31 6.20 5.22 3.57 6.05 Rural 6.79 5.56 9.33 - 6.79 Child mortality rateh All women Total 0.18 0.08 0.09 0.00 0.16 Urban 0.13 0.04 0.08 0.00 0.11 Rural 0.20 0.14 0.12 0.00 0.20 Women age 35-49 Total 0.18 0.12 0.05 0.00 0.17 Urban 0.13 0.16 0.05 0.00 0.11 Rural 0.20 (.09 0.00 - 0.19 - Not available. Note: Working samples are 1,943 women for C6te d'lvoire and 2,237 for Ghana. a. Number of children born alive. b. Number of deaths of children under age five per live birth. Source: Benefo and Schultz (1994, tables 3 and 4). Benefo and Schultz 137 older women's. The marked divergence in age patterns suggests that health condi- tions may have worsened in Ghana or at least not improved in the 1970s and early 1980s. Similar levels of child mortality are currently reported for the rural populations of both countries, although by the 1980s the urban population of C6te d'lvoire appeared to have achieved a lower level of child mortality than that of Ghana. III. ESTIMATES OF ALTERNATIVE SPECIFICATIONS Models of fertility determinants are estimated on the basis of three specifica- tions. For model I, child mortality is assumed to be exogenous and the fertility equation can then be estimated by OLS. For model II, child mortality is endogenous, and is identified by instruments measuring community health ser- vices and environment. Because household composition, husband characteris- tics, and household income are likely to be endogenous and cannot be readily identified, these intervening variables are implicitly solved out of the model and thus omitted from the estimated model (Schultz 1994). For model III, household composition and husband characteristics are assumed to be exogenous and po- tentially affect child mortality and fertility. Child mortality is otherwise det- ermined and estimated as in model ll. On conceptual grounds, model ll is preferred, but empirically the identifica- tion of child mortality may not be satisfactory. Model 1, which is an unstruc- tured description of the data, may then provide more reliable estimates, although potentially they are biased due to simultaneity and measurement error. Model III is reported to assess the robustness of the structural model It estimates by conditioning both equations on household composition and husband character- istics despite their potential endogeneity. Child Mortality Determinants Multivariate results are presented in table 2, where the mortality rate of chil- dren through the first five years of life is estimated using OLS in regression 1, based on model I (or 11). That is, the effects of household income, husband characteristics, and family headship are assumed endogenous and implicitly solved out of these reduced-form estimates. In regression 2, based on model lll, hus- band characteristics and family headship are assumed exogenous and are included. Mother's education, as shown in table 1, is associated with lower child mor- tality when the population is cross-tabulated by age and rural or urban resi- dence, a pattern observed in many studies of African demographic surveys (for example, Cochrane, O'Hara, and Leslie 1970; Frank and Dakuyo 198S; Aly and Grabowski 1990; Okojie 1991; Maglad 1994). The results show that when con- trols are included for the community health infrastructure, health problems, food prices, household assets, ethnicity, and region, the mother's education is more weakly-although still negatively-related to child mortality (table 2). 1.38 THF WORLD BANK LCoNOMI( RkFVIF\-, VOl. 1. NO. I The effects of education by level of schooling are not individually statistically significant. Each year of secondary schooling completed by the mother in C6te d'lvoire and each year of middle school completed by the mother in Ghana is associated with a reduction in her child mortality rate of about 0.01. The mother's height, an indicator of her health human capital, is statistically significant and negatively associated with child mortality in Ghana. Household assets per adult are associated with lower child mortality in Ghana, but not in C6te d'lvoire. We interpret these patterns as suggesting that access to economic resources in Ghana has a larger effect on child mortality than it does in C6te d'lvoire (Benefo and Schultz 1 994). The west forest region of C6te d'lvoire has notably higher levels of child mortality than Abidjan, the excluded region. All regions in Ghana except urban forest and rural savannah have higher child mortality rates than Accra. Ethnic differences in child mortality are statistically significant as a group in both coun- tries, with the Krou in C.ote d'lvoire reporting lower mortality compared with the Akan (omitted), and the Ewe and Ga-Adangbe with lower mortality in Ghana. Christian women also report lower mortality in Ghana. As a set, the comiimunity variables do not explain a great deal of variation in child mortality anid consequently provide a weak basis for identifying the exogenous effects of child mortality in the subsequent fertility equation. None- theless, these instruments are statistically significant as a group, a common criterion for reliable IV estimates (Bound, Jaeger, and Baker 1995).4 Living farther from a market is associated with higher child mortality in Ghana, and distance to a clinic is not statistically significant. Conversely, in C6te d'lvoire, womein who reside closer to a clinic experience lower child mortal- ity, but the effect of distance to a market has an unexpected sign. Greater rainfall reduces child mortality in both Ghana and C6te d'lvoire, but the negative effect is not statistically significant in either country. Communities reporting a serious health problem with malaria and measles in C6te d'lvoire show higher child mortality, but a local problem with diarrhea is related to the poorer survival prospect of children in Ghana. Water and sanitation do not exhibit any noted relationship with clild mortality at the community level, although later analysis of these variables suggests benefits may differ across groups of mothers. The food prices are jointly significant, but they are difficult to evaluate on an individual basis for reasons noted earlier. Adding control variables for household composition and husband characteristics in regression 2 does not improve the estimates significantlv or alter substan- tially the other relationships discussed. 4. T-he joint F te.t on the exclusioll of the set of communllity health and price variables from the child miortality equation including the predicted income variable based on model 11 is F = 2.61 (15,1904) in C6te d'Ivoire, and 1 = L3.06 ( 1 7,219 1 ) in Ghana. Both are statistically significant at p< 0.002. Including household composition, and husband characteristics increases slightly these F tests. There is always the possibility that the communinty variables also affect fertility, and thev Would then he invalid instruments for identifying child mortalirv in the fertilitn equation. Benefo and Schultz 139 Table 2. Regressions on the Child Mortality Rate for C6te d'lvoire, 1985-87, and Ghana, 1987-89 Cote d'lvoire Ghana Regres- Regres- Regres- Regres- Variable sion I sion 2 sion I sion 2 Individual variables Woman's schooling (years) Primary (1-6) -0.0035 -0.0022 -0.0015 -0.0014 (1.01) (0.62) (0.47) (0.42) Middle (1-4) -0.0006 -0.0000 -0.0092 -0.0089 (0.07) (0.00) (1.80) (1.73) Secondary and higher education -0.0109 -0.0115 -0.0053 0.0058 (1-3 or more) (1.18) (1.21) (1.24) (1.34) Woman's height (In meters) -0.111 -0.101 -0.330 -0.352 (0.78) (0.70) (2.43) (2.58) Household assets (local currency), 0.0182 0.0190 -0.0227 -0.0239 (0.62) (0.64) (1.46) (1.54) Woman's ageh 25-29 0.0001 0.0008 0.0038 0.0088 (0.01) (0.04) (0.18) (0.40) 30-34 0.0108 0.0109 0.0080 0.0143 (0.51) (0.49) (0.38) (0.62) 35-39 0.0192 0.0183 -0.0103 -0.0021 (0.86) (0.75) (0.46) (0.08) 40-49 0.0280 0.0273 -0.0166 -0.0119 (1.36) (1.17) (0.76) (0.48) 50 or more 0.0671 0.0656 -0.0258 -0.0243 (3.24) (2.73) (0.69) (0.62) Household composition and husband characteristics' No husband present (dummy) 0.0114 0.154 (0.10) (1.17) Woman head of household (dummy) 0.0013 -0.0233 (0.06) (1.07) Husband's schooling (years) Primary (1-6) 0.000 -0.0013 (0.02) (0.26) Middle (1-4) -0.0095 -0.0021 (1.30) (0.29) Secondary and higher education (1-3 or more) 0.0029 -0.0025 (0.41) (0.75) Husband's height (In meters) 0.0937 0.418 (0.62) (2.47) Husband's age (years) -0.0014 0.0037 (0.40) (0.95) Husband's age squared (xl 0-2) 0.0014 0.0043 (0.45) (1.05) Woman's current residence (dummy variables)d.e Other urban/urban coast 0.024 0.024 0.064 0.061 (1.00) (1.01) (2.51) (2.39) East forest/rural coast 0.043 0.043 0.058 0.052 (1.47) (1.42) (1.66) (1.50) (Table continues on the following page.) 140 THE WORLD) BANK ECONOCMIC REVIEW, VOL. 10, NO. I Table 2. (continued) C6te d'lvoire Ghana Regres- Regres- Regres- Regres- Variable sion I sion 2 sion I sion 2 West forest/urban forest 0.092 0.091 0.030 0.026 (2.60) (2.53) (0.90) (0.77) Savannah/rural forest 0.027 0.026 0.073 0.070 (0.791 (0.77) (2.00) (1.92) n.a./urban savanmah 0.059 0.070 (1.62) (1.92) n.a./rural savannah 0.062 0.052 (1.58) (1.41! Womran's ethnicity or language (dummy variables)di Other/Ewe 0.018 0.0116 -0.070 -0.072 (0.97) (0.61) (4.24) (4.33) Krou/Ga-Adangbe -0.049 -0.050 -0.079 -0.084 (2.23) (2.22) (3.45) (3.61) Mande-North/Dagbani -0.013 -0.017 0.022 0.015 (0.60)) (0.78) (0.63) (0.43) Mande-South/Hausa 0.017 0.016 -0.032 -0.038 (0.84) (0.79) (0.82) (0.97) Voltaic/Nzema -0.024 -0.028 -0.000 0.002 (0.99) (1.16) (0.01) (0.03) Other language -0.022 -0.027 (1.21) (1.45) Wonomna's religion (dummy uariables)g Muslim -0.0033 -0.0048 (0.11) (0.16) Christian -0.0596 -0.0567 (2.45) (2.32) Traditional and other 0.0032 0.0003 (0.12) (0.01) Community variables Proportion of cluster sample households -0.019 -0.019 -0.0044 -0.0009 with toilet or latrine (0.83) (0.78) (0.23) (0.05) Proportion of cluster sample households 0.0088 0.0102 -0.0022 -0.0011 with protected water source, including (0.52) (0.60) (0.13) (0.06) piped water and wells with pumps One of the two most serious community health problems (dummy variables) Malaria 0.0)42 0.042 -0.0083 -0.0079 (2.20) (2.17) (0.46) (0.44) Diarrhea 0.021 0.021 0.0336 0.0333 (1.26) (1.27) (2.00) (1.98) Measles, chicken pox, or other 0.036 0.036 -0.0097 -0.0079 infectious illnesses (2.11) (2.12) (0.63) (0.51) Distance to nearest health clinic (kilo- 0.0011 0.0011 0.0007 -0.0006 meters for C6te d'lvoire, miles for Ghana) (2.97) (2.91) (0.64) (0.57) Distance to nearest marketplace (kilo- -0.0024 -0.0025 0.0025 0.0026 meters for C6te d'lvoire, miles for Ghana) (1.59) (1.66) (2.69) (2.70) Benefo and Schultz 141 Table 2. (continued) Cote d'lvoire Gbana Regres- Regres- Regres- Regres- Variable sion I sion 2 sion I sion 2 Child immunization campaign in last -0.0084 -0.0077 five years (dummy) (0.45) (0.40) Public health expenditures per person in -0.361 -0.303 the province (x103, 1988 cedis) (0.48) (0.40) Rainfall (centimeters per year)' -0.0006 -0.0005 -0.0011 -0.0010 (1.27) (1.18) (1.40) (1.29) Share of tree crops' -0.0208 -0.0232 -0.0623 -0.0635 (0.84) (0.93) (2.04) (2.08) Prices in the communityd'l Manioc/cassava 0.0224 0.0240 0.0030 0.0029 (0.15) (0.16) (2.29) (2.20) Bananas/maize -0.0599 -0.0641 0.0023 0.0023 (0.41) (0.43) (2.98) (2.97) Fish/fish -0.0756 -0.0744 -0.0000 -0.0000 (2.11) (2.07) (0.54) (0.64) Beef/eggs -0.118 -0.120 -0.0074 -0.0075 (1.85) (1.87) (3.52) (3.56) Palm oil/antibiotics -0.0330 -0.0321 0.0214 0.0218 (1.82) (1.77) (1.88) (1.92) Peanut butter/sugar -0.0723 -0.0706 0.0012 0.0011 (1.82) (1.77) (3.52) (3.46) Round 2 (dummy) 0.004 0.0004 -0.001 -0.001 (0.21) (0.02) (0.11) (0.08) Round 3 (dummy) -0.017 -0.021 (0.78) (0.96) Intercept 0.396 0.379 0.135 0.0119 (3.64) (2.56) (0.94) (0.06) R2 0.087 0.088 0.098 0.102 n.a. Not applicable. Note: The dependent variable is the child mortality rate. The coefficients were estimated using ordinary least squares. The absolute values of the t-statistics are in parentheses. The sample size is 1,943 in C6te d'lvoire and 2, 237 in Ghana. a. The value of the household's assets includes the value of owned land, business assets, and ten times other property income per year per adult in thousands of CFA francs for C6te d'lvoire and thousands of cedis for Ghana. b. The excluded category is women age fifteen to twenty-four. c. The sample average is based on a variable that is set to zero for women who have no reported husband. For example, the average number of years of husband's primary education for women with husbands in C6te d'lvoire is (1.31)/(1-0.240) = 1.72 years. d. The term before the slash (/) applies to C6te d'lvoire, the term after applies to Ghana. e. The excluded category is Abidjan or Accra. f. The excluded category is Akan. g. The excluded category is no response. h. Rainfall in the area, or at the nearest weather station, in annual average centimeters per year in Ghana and in centimeters for the previous year in C6te d'lvoire. i. Proportion of cluster sample household land area that is farmed in tree crops such as cocoa, coffee, bananas, or coconuts. j. Local prices are averaged and adjusted for inflation to the initial survey date in CFA francs for C6te d'lvoire and in cedis for Ghana. Source: Authors' calculations. 142 111F WX'oRI I! BANK F( ()\()\II R I \ IlKA, Vodl . 1NW. I As noted in the cross-tabulations, older mothers experience higher child mor- talitv rates even when they have the same education and reside in similar com- munities in Cote d'lvoire, suggestling a secular monotonic improvemenit in child mortality in that country. The same pattern of age coefficients is not observed in Ghana, and although the age coefficients are not statistically significant, child mortality ainonig mother-s age twenty-five to thirty-four may be somewhat higher than that amonig older mothiers. Fertility Determininzzts Women's edUcation is associated with lower fertility in table 3, as seen in virtually all studies of fertility (Schultz 198 ). As others have noted in some regions of Africa, hiowever, the effects become statistically significant only after the completion of the first few years of primary schooling (Ainsworth, Beegle, anid Nyarmete in this issue). This coLlld be partially explained by the low wage returnis oin primary education, particU larly in Ghana, or the low content and quality of early schoolilg. A maj-or dissimilarity between the two countries emerges in the effect of other economic endowment variables on fertility. Houselhold assets per adult are positively related to fertility in CUte d'lvoire and negatively related in Ghana. The magniltude and statistical significance of the asset effect on fertility depends on whether it is estimated in regression 2 or in regressioni 3, where huslband characteristics are con- trolled. The mother's height is interpreted here as a measure of her health status and produCti vity. It is positively related to her fertility in Cote d'lvoire and has no relation in Ghana. As in a previous study (Benefo and Schultz 1 994) that estimated houselhold income and instrumented for it in the fertil- ity equation, we interpret these results as suggesting that income from male workers and nonLhunian capital appear to he associated with greater fertility in C6te d'lvoire but iiot in Gliaia. The effect of chilld mortality on fertility is statistically significant only if child mortality is assumiled exogenous, as in model 1. When child mortality is treated as endogenous and identified by community variables representing health ser- vices, conditionis, and food prices, the fertilitv response to the predicted mortal- ity variable is niot statistically significant in either country, and even changes sign in Ghana.i This flindinig is in contrast with maniy other studies using a va- riety of methodologies (Taylor, Newman, and Kelly 1 976; Cantrelle, Ferry, and Mondot 1978; Olsen 1 980; Lee and Schultz 1982; Rosenzweig and Schultz 1982; Cochrane and Zachariah 1983; Oko;ie 1 991; Maglad 1994). The distance to the mnarket is again an important factor in Ghana, where it predicts higher levels of fertility, as it did of chiild mortality. Religion is not sig- nificantly relatedl to fertility in Ghanla, given the other variables in the regression. Regiolis are impportanit in both countries, although ethnic categories are more significant as a groip in Ghana thani in Cote d'lvoire. Husbland characteristics i. Chid Illlrtrlit\ is accepted as exugClluLls according to tie lHatusmian test ill both C,hana and Ckte d'lvoire at the 10) pero-cot coeiicduncc level. Benefo and Schultz 143 Table 3. Regressions on the Number of Children Born to Women, Cote d'Ivoire, 1985-87, and Ghana, 1987-89 Cote d'lvoire Gbhna Regres- Regres- Regres- Regres- Regres- Regres- Variable sion I sion 2 sion 3 sion I sion 2 sion 3 Individual variables Woman's schooling (sears) Primary (1-6) 0.0133 0.0131 0.0364 -0.021 -0.024 -0.011 (0.35) (0.34) (0.91) (0.85) (0.95) (0.45) Middle (1-4) -0.164 -0.164 -0.125 -0.169 -0.185 -0.169 (1.84) (1.83) (1.35) (4.31) (4.51) (4.13) Secondary and higher -0.231 -0.232 -0.218 -0.133 -0.122 -0.106 education (1-3 or more) (2.30) (2.25) (2.(9) (4.06) (3.66) (3.16) Woman's height (In meters) 3.40 3.39 2.66 -0.064 -0.532 -0.552 (2.20) 12.16) (1.2) (0.06) (0.48) (0.51) Child mortality rate' 1.07 1.01 1.31 0.480 -1.10 -1.14 (4.30) (0.51) (0.68) (2.95) (0.93) (0.98) Household assets (local currency)( 0.538 (0.539 0.393 -0.0473 -0.0797 -0.10)6 (1.68) (1.67) (1.24) (0.40) (0.65) (0.88) Woman's age, 25-29 0.902 0.903 0.092 0.878 0.886 0.749 (4.03) (4.01) (3.11) (5.32) (5.36) (4.51) 30-34 1.91 1.92 1.51 1.97 1.99 1.71 (8.30) (8.21) (6.36) (12.0) (12.0) (9.84) 35-39 3.21 3.21 2.-5 2.85 2.83 2.47 (13.2) (12.9) (10.5) (16.3) (16.1) (13.0) 40-49 3.59 3.59 3.06 4.33 4.31 3.99 (16.0) (15.3) (11.9) (26.0) (25.6) (21.2) 50 or more 3.41 3.41 3.27 4.32 4.28 4.08 (15.1) (13.0) (11.3) (15.1) (14.8) (13.7) Household composition and busband characteristics' Nohusband present (dummy) 1.16 2.15 (0.951 (2.18) Woman head of household (dummy) -0.066 0.050 (0.2-t) (0.30) Husband's schooling (years) Primary (1-6) 0.082 0.017 (2.28) (0.45) Middle (1-4) -0.030 -0.071 (0.39) (1.31) Secondary and higher education (1-3 or more) -0.036 -0.001 (0.49) (0.02) Husband's height (In meters) -4.49 -1.16 (2.80) (0.91) Husband's age (years) 0.174 0.146 (4.84) (4.88) Husband's age squared (x102) -0.158 -0.138 (4.68) (4.46) (Table continues on the follouwing page.) 144 FHF WORLD BANK ECONO.MIC REVIEW. VOL. 10, NO. I Table 3. (continued) C6te d'lvoire Ghana Regres- Regres- Regres- Regres- Regres- Regres- Variable siotn I sion 2 sion 3 siran I sion 2 sion .3 Wonman's current residence (dunmnv uvariables)/f Other urban/urban coast 0.272 0.273 0.424 0.422 0.532 0.587 1.17) (1.16) {1.84) (2.25) (2.60) (2.91) East forest/rural coast 0.751 0.754 0.8(0 0.834 0.880 0.932 (3.05) (2.70) (2.92) (3.58) (3.73) (4.01) West forest/urban forest 0.501 0.506 0.627 0.268 0.311 0.395 (1.96) (1.54) (1.94) (1.18) (1.35) (1.74) Savaninah/rural forest 0.381 0.385 0.337 0.578 0.679 0.675 (1.37) (1.17) (1.04) (2.33) (2.62) (2.63) n.a./urbani savannah 0.666 0.746 0.789 (2.65) (2.86) (3.07) n.a./rural savannah 0.249 0.349 0.344 (I .()9) (1.45) (1.45) Woman s etbnicity oir language (diminv variables) , Other/Ewe -0.215 -.2 14 -0.301 -0.346 -0.456 -0.476 (1.09) (1.06) (1.47) (2.86) (3.11) (3.30) Krou/Ga-Adangbe -0.25(0 -0.252 -0.419 -0.5.53 -0.688 -(0.707 (1.()9) (1.03) (1.73) (3.18) (3.42) (3.57) Mande-North/Dagbanii -0.196 -0.196 -0.2-4 -((.622 -0.576 -0.872 (0.87) (0.87) (1.22) (2.33) (2.14) (3.19) Mande-South/Hausa -0.261 -0.260 -0.500 0.202 0.176 0.009 (1.26) (1 .22i (2.38) (0.68) (0.59) (0.03) Volraic/Nzema -0.605 -0.604 -0.575 -0.640 -0.637 -0.442 (2.69) (2.66) (2.54) (1.50) (1.50) (1.03) Other language -0.080 -0.102 -0.274 (0.59i (0.75) (1.96) Womnan's religion (dummvy lariables)h Muslim 0.204 0.231 0.252 (().9 1) (1.02) (1.13) Christian -0.005 -0.076 0.011 ((0.03) (0.39) ((.06) Traditionial and other 0.145 0.185 0.2 12 ((.7() (0.881 (1.03) Community variables Distance to nearest mnarketplace -0.0027 -0.0028 0.0046 0.0)140 0.0170 0.0188 (kilometers for C6te d'lvoire, (0.17) (0.18) (0.30) (2.10)) (2.41) ().71) miles for Ghana) Rainfall (centimeters per year)' -0.0054 -0.0054 -).0028 0.0064 0.0041 0.0038 (1.23) (1.19) (0.63) (1.14) (0.69) (0.65) Share of tree crops 0.216 0.215 0.087 0.0077 -0.079 -0.105 (0.85) (0.51) ((0.33) (0.33) (0.34) (0.46) Round 2 (dummy) 0.141 0.142 0.1()7 -0.085 -0.088 -(0.089 (0.64) (0.63) (11.48) (1.(5) (1.08) (1.12) Benefo and Schultz 14 5 Table 3. (continzued) Cote di'lvuire Ghana Regres- Regres- Regres- Regres- Regres- Regres- Variable sion I soun 2 sin 3 ? Sfn I sioni 2 stoI 3 Round 3 (dummy) -.049 -0.049 -0.085 (0.21) (0.21) (0.38) Intercept 1.64 1.65 0.0060 2.04 2.66 0.220 (1.74) (1.61) (0.00) (3.61) (3.63) (0.19) R' 0.268 0.261 0.299 0.431 0.429 0.451 Hausman tests of child mortality 0.027 -0.145 1.34 1.43 being "exogenous" n.a. Not applicable. Note: The dependent variable is number of children born alive to women over age fifteen. For regression 1, child mortality is assumed to be exogenous, and the fertility equation is estimated using ordinary least squares. Regression 2 is estimated using instrumental variables (IV); child mortality is endogenous and identified by instrimetits measuring cominuitity health services and enivironment. Regression 3 is estimated using IV; household composition and hiusband characteristics are assUrmed to be exogenous; child molrtality is endogenous and identified as in regression 2. The absolute value of the asymptotic t-statistic is reported in parentheses. The sample size is 1,943 in Cote d'lvoire and 2,237 in (Ghan.a. a. Number of deaths of childreni under age five per live birth. b. The value of the household's assets includes the value of owvned land, business assets, and ten tinies other property income per year per adult in thousands of (i,\ francs for Cote d'lvoire and thousands of cedis for Ghana. c. The excluded category is womeni age fifteen to tsventy-foir. d. The sample average is based on a variable that is set to zero for women who have no reported husbanid. For example, the average number of years of hUsband'S primary education for woomen with husbands in C6te d'lvoire is ( 1.31)/(1-0.240) = I .72 years. e. The term before the slash (/) applies to Cote d'lvoire; the term after applies to Ghana. f. The excluded category is Abidjan or Accra. g. The excluded category is Akan. h. The excluded category is no response. i. Rainfall in the area, or at the nearest weather stationi, in annIal average centimeters per year in Ghana and in centimeters for the previous year in Cote d'lvoire. j. Proportion of cluster sample hoLIsehold land area that Is farmed in tree crops such as cocoa, coffee. baisanas, or coconuts. Sou rce: Authiors' calculations. and family composition variables are not jointly significant as a set of variables in regression 3 for understanding the cumulative patterns of fertility. The differ- ential effect of male and female education on fertility has been observed widely (Ainsworth, Beegle, and Nyamete in this issue). Increases in middle and second- ary education for women in either country are associated with substantial de- creases in fertility, but advances in male education are not associated with sig- nificant declines in fertility in either country. These results do not, however, give support to Caldwell and Caldwell's (1987) argument that fertility is determined in a different cultural context in Africa. The different effect of male and female education on fertility is a common conclusion drawn from many economic models and from empirical studies in Africa and elsewhere (Willis 1974; Schultz 1981 ). 146 THE WORLL) BANK ECONOMIC REVIEW, VOL. 10, NO. I In sum, we find different patterns of fertility for the two countries. In C6te d'Ivoire, assets are weakly associated with higher levels of fertility. Economic growth associated with the formation of nonhuman capital may be associated with small declines in fertility in Ghana, but may not have this immediate con- sequence in C6te d'lvoire. Child mortality reductions in both countries are asso- ciated with fertility declines, but the individual association in model I suggests that only about one-sixth of any decline in child mortality will be translated into an offsetting fertility decline in C6te d'lvoire (the comparative figure for Ghana is one-twelfth). These are smaller responses of fertility to child mortality than others have found from individual data, where child mortality is assumed ex- ogenous (Cochrane, O'Hara, and Leslie 1970; Rosenzweig and Schultz 1983). This result may be caused by the increased number of control variables included in this study or by a reduced responsiveness in fertility when the overall level of child mortality is as high as in these West African countries. Women's education in both countries, particularly schooling beyond the pri- mary level, is linked in this and other studies to fertility declines of a substantial magnitude-each additional year of education for women is associated with their having 0.1 to 0.2 fewer births. By contrast, advances in the educational attainment of men will yield few such dividends in terms of slowing population growth. Social investment in women's education is economically productive both in the labor force, and as a source of demographic externalities that contribute to slowing the rate of population growth. Fertilitv Estimates by Age and Rural-Urban Residence The fertility equation is reestimated in table 4, for women age twenty-five to thirty-four and thirty-five to forty-nine living in rural and urban areas, treating child mortality as exogenous. These subsamples are about one-fifth the size of those in the overall regressions, and consequently the estimated coefficients are subject to much more sampling variability. However, number of children born at these later ages is a better proxy for lifetime fertility, and the importance of the different variables for fertility may differ between rural and urban areas. Education of women, when it is statistically significant, always has a negative effect on fertility. The coefficient on the levels of education vacillates consid- erably in the small subsamples, to the point where it is undefined at two levels for rural Ivorian women because in the sample there are no rural women thirty- five to forty-nine with schooling beyond the primary level. Urban women age twenty-five to thirty-four exhibit the clearest evidence that an additional year of postprimary schooling is associated with a reduction in fertility of 0.2 children in C6te d'lvoire and about 0. I children in Ghana. The estimates for these educa- tion coefficients do not change greatly for older women, but the precision of the estimates decreases. The positive effect of household assets on fertility is evident only among younger rural women in C6te d'lvoire. The negative effect of assets on fertility may be evident among older groups of Ghanaian women in both rural and urban Table 4. Regressions on the Number of Children Born to Women Age 25-34 and 35-49, by Rural and Urban Residence, C6te d'lvoire, 1985-87, and Ghana, 1987-89 C6te d'Ivoire Ghana Rural Urban Rural Urban Individual or household variable 25-34 .35-49 25-34 35-49 25-34 35-49 25-34 35-39 Woman's schooling (years) Primary (1-6) 0.0656 0.0226 0.0363 -0.129 -0.0331 0.0704 -0.0143 -0.0460 (0.97) (0.11) (0.82) (1.00) (0.76) (0.71) (0.41) (0.72) Middle (1-4) -1.02 n.a. -0.203 -0.069 -0.185 -0.104 -0.175 -0.204 (0.91) (2.38) (0.26) (2.60) (0.58) (3.60) (1.98) Secondary and higher education 0.243 n.a. -0.235 -0.163 -0.155 -0.145 -0.0892 -0.136 (1-3 or more) (0.53) (2.30) (0.96) (1.18) (0.96) (2.16) (2.40) Woman's height (In meters) 0.0475 5.76 1.50 3.77 -1.39 1.65 -0.862 2.92 (0.02) (1.63) (0.55) (0.62) (0.70) (0.49) (0.60) (1.10) Household assets (local currency):' 0.75 -0.52 1.60 -10.7 -0.146 -1.30 0.086 -0.303 (2.03) (0.12) (1.(0) (1.25) (0.42) (1.12) (0.77) (0.98) Child mortaliry rate (exogenous)b 0.487 0.986 0.826 0.562 0.175 0.952 0.534 0.722 (1.18) (1.66) (1.69) (0.55) (0.70) (1.48) (2.41) (1.36) Woman's age (years) 0.211 0.0107 0.232 (.099 0.239 0.204 0.203 0.148 (5.27) (0.33) (6.58) (2.16) (9.10) (6.71) (9.52) (6.41) Distance to nearest marketplace 0.0159 -0.0289 0.0236 -1.07 0.0122 0.0275 0.0073 0.0151 (kilometers for C6te d'lvoire, (0.69) (0.89) (0.04) (2.00) (1.34) (1.44) (0.59) (0.78) miles for Ghana) Rainfall (centimeters per year)c -0.0056 -0.0154 -0.0007 -0.0172 0.0147 -0.0131 0.0011 0.0265 (0.78) (1.69) (0.09) (0.93) (1.59) (0.71) (0.14) (1.85) R2 0.143 0.079 0.218 0.139 0.252 0.199 0.227 0.263 Sample size 299 428 343 224 436 344 663 560 n.a. Not applicable. No women fit this category. Note: The dependenit variable is the number of children born alive to women in the age categories. Regressions are estimated using ordinary least squares. Child mortality rate is treated as exogenous. Absoliue value of the t-statistic is reported in parentheses. Controls are also included for ethnic or religion groups and survey round. a. The value of the household's assets includes the value of owned land, business assets, and ten times other property income per year per adult in thousands of CFA francs for Cote d'lvoire and thousands of cedis for Ghana. b. Number of deaths of children under age five per live birth. c. Rainfall in the area, or at the nearest weather station, in aniniual average centimileters per year in Ghana and in centimeters for the previous year in Cote d'lvoire. Source: Authors' calculations. 148 THF. WORI D BANK ECONOMI(C REVIEW, VOl. 10, NO. I areas. Expanding female education is associated with decreases in fertility in both countries, but the scarcity of educated women resident in rural C6te d'lvoire makes it difficult to forecast whether this process lowers fertility among those staying in rural areas or only contributes to the decline in national fertility be- cause it contributes to rural-urban migration. In the rural and urban subsamples, child mortality is again positively asso- ciated with fertilitv at all age subsamples. Child mortality is statistically signifi- cant in the older rural and younger urban samples in C6te d'lvoire and in the older rural and both urban samples in Ghana. The implied replacement rate for births to a child death by the time a woman has completed her childbearing at age thirty-five to forty-nine is 0.20 to 0.25, somewhat higher than previously estimated across all ages for Ghana in table 3. Overall, the disaggregation of the sample by age and rural or urban areas gives us more confidence in the findings based on the age-aggregated fertility functions, although the much smaller samples prevellt US from obtaining precise estimates. Who Benefits Most from Local Programs? Some studies of fertility and child mortality have found evidence that the benefits which a population receives from community health and sanitation ser- vices depend on the level of women's education. Education allows women to reduce their children's mortality and reduce unwanted childbearing (Schultz 1981, 1988a, 1988b, 1992; Rosenzweig and Schultz 1982; Rosenzweig and Wolpin 1986; Barrera 1990, 1991). The empirical pattern observed most often in this literature suggests that higher levels of these community health services are as- sociated with greater benefits for less-educated women and their families. Fe- male education and these community health and sanitation services are "substi- tutes" in the production of child health, and these programs reduce educational differences in health outcomes. The opposite pattern is also noted, but less of- ten, where a particular social service, such as piped water, is found to be more effective in the hands of the better educated, in which case these services "com- plement" the mother's educational attainment. To test the hypothesis that community programs in Ghana and C6te d'lvoire differentially benefit women who have different levels of education, explana- tory variables are defined as the product of the mother's years of completed education and the four community characteristics. These interaction variables are included in the regressions reported in table 5, explaining child mortality and fertility that also control for (overall) years of education, age, and com- munity variables. In the Ghana child mortality regression, the negative coefficient on the interaction between schooling and protected water indicates that the avail- ability of protected water sources in the community lowers child mortality by a greater amount among more-educated mothers. Maternal education also has a stronger effect in reducing child mortality among Ghanaian moth- ers in rural areas than among those in urban areas. These estimates imply Benzefo and Schultz 149 Table 5. Child Mortality and Fertility Regressions with Interactions betweenl Mother's Education and Community Characteristics, C6te d'Ivoire, 1985-87, and Ghana, 1987-89 C6te d'tvoire Ghbana Child Child mortalitv mizortality Selected explanatory variable (age 0-4) Fertility (age 0-4) Fertility Woman's schooling (years) -0.034 0.0096 -0.0031 0.0534 (2.84) (0). 07) (1.16) (2.60) Woman's height (In meters) -0.0873 2.94 -0.258 -0.208 (0.63) (1.92) (1.89), (o.20) Rural residence (durnmm) 0.027 0.861 0.063 -0.0)42 (1.44) (4.14) (3.91) (0.34) Distance to nearest health clinic 0.0011 0.00)87 -0.00)038 (0.0162 (kilometers for Cbte d'lvoire, (3.25) (2.30) (0.34) 1.87) miles for Ghanai Proportion of cluster sample -0.013 I.017 0.0682 0.0073 households with protected water (0.76) ((M.10) (3.47) (0.05) source, including piped water and wells with pumps Proportion of cluster sample 0.041 0.375 -0.0649 ).115 households with toilet or latrine (1.91) 1.60) (3.31) (0.-77) Schooling interacted with Rural residence 0.0089 0.0464 -0.013 0.()1 (0.82) 0.39) (4.33) (0.75) Distance to nearest health clinic 0.0((1 -().00(11 (0.00043 0.0(1069 (0.24) (0.21) (1. 79) ((.38) Proportion of cluster sample -0.0032 -0.0592 -().0109 -0.0632 households with protected water (0.61) (1.03) (3.74) (2.85) source, includinig piped water and wells with pumps Proportion of cluster sample 0.0336 -0.0394 0.0057 -0.035 households with toilet or latrinie (2.88) ((1.31) (1.70) (1.39) R` 10.071 0.251 0.048 0.419 Note: Regression estimates are hased on a reduced-torm specification. Controls also included for age dummies, survey round, rainfall, and distance to miarket. t-statistics aLre in parenithieses. The sample size is 1,943 in Cote d'voire and 2,2 7 in Ghana. Source: Authors' calcul)1tion1s. that public sector subsidies to increase female education in Ghana would have a greater effect in reducing child mortality if they were allocated to rural rather than to urban areas. Improvements in water supplies are more effective in reducing child mortality if they are provided to communities where more women are better educated, possibly because education teaches moth- ers how to use water for hygienic purposes. The remoteness of a clinic in Ghana raises child mortality by a greater amount for more-educated women than for less-educated. Thus, the proximity to a clinic appears to "com- plement" women's education. 150 THE WORLD BANK ECONOMIC. ROVIEW. VOL. 10, NO. I The prevalence of modern sanitation facilities in the form of toilets and la- trines in the community is the only community variable that appears to interact with maternal education in determining child mortality in C6te d'lvoire, where it complements a mother's education. Thus the availability of toilets in the local community appears to increase educational differentials in child mortality in both countries. The sparseness of measured educational interactions in C6te d'lvoire com- pared with Ghana may be partly due to the lower levels of education among older Ivorian women. When the maternal education interactions are disaggre- gated by primary, middle, and secondary and higher education, as in the earlier regressions, several additional regularities are evident. The interaction between maternal education and the frequency of toilet or latrine availability in the clus- ter is stronger for primary schooling, the only interaction variable with the level of women's education that is separately an important correlate with child mor- tality in C6te d'Ivoire. In Ghana these child health benefits from the community sanitation practices improve for mothers with primary, middle, or secondary education. In Ghana, primary and secondary education of mothers appear to protect their children from the health disadvantages of rural residence and to strengthen the benefits they realize from community protected water supplies. IV. CONCLUSIONS AND RESEARCH PRIORITIES To assess how development and social welfare programs affect such outcomes as child mortality and fertility, the researcher must measure the critical dimen- sions of these policies and programs. Moreover, the variation observed across a surveyed population in these policies and programs must also be independent of the host of confounding factors that might otherwise explain these outcomes. More research is needed to quantify independent variation in the character and quality of local health care that is produced by specific government policies, programs, and the pricing of services. In the collection of policy-oriented household surveys, such as the LSMS program, the information gathered at the community level is critical for linking policy interventions to socioeconomic outcomes. Relatively little is known about the accuracy or relevance of current responses to community questionnaires regarding the availability of health, edu- cation, or family planning services, or how the measured availability of services actually affects the welfare of the neighboring population. The design, valida- tion, and refinement of community policy questionnaires that parallel household sample surveys are neglected topics for research. Community policy question- naires could be important for improving development welfare policies. Are community health programs allocated in response to prior health condi- tions across communities, or do these program allocations affect the migration of people with special health needs or preferences? If either of these processes occurs, cross-sectional relations between programs and outcomes, even when programs are suitably measured, can yield biased estimates of program effects. Benefo anid Schultz 151 The only way to be confident that cross-sectional variation in community pro- grams is independent of unobserved community conditions is to design the pro- grams to achieve a phased sequence of program interventions that are orthogonal to such unobserved factors. Unless the variation across women in their child mortality rates is then explainable by community health programs and environ- ments, we should be agnostic about the capacity of the public health sector, as it is currently measured, to improve substantially child survival. By the same to- ken, if education and family planning programs cannot be shown to enhance significantly school enrollments and achievements and to reduce unwanted fertil- ity, there should be skepticism regarding the effectiveness of expanding these existing social welfare systems. We found indications in Ghana that economic resources of households, ma- ternal education, access to markets, and food prices are all associated with child mortality. Residence closer to a health clinic (public or private) is not a good predictor of child mortality in Ghana, perhaps because local proximity to a clinic does not capture the effect of the prices for, or qualitv of, clinic-provided health care. In both countries, sanitation infrastructure, in the form of com- munity toilets and latrines, may slightly increase child survival prospects, but only for children of less-educated mothers. The higher levels of child mortality in rural areas of Ghana are less severe for better-educated mothers. Conversely, the health advantages of urban residence are particularly beneficial for the chil- dren of the uneducated mothers (table 5). In contrast, communities in Ghana in which a larger fraction of the sampled cluster can rely on protected water sup- plies do not report decreased child mortality, except perhaps for women with more than six years of education (table 5). Education may teach women how to use improved water supplies effectively for reducing health risks for their children. In C6te d'lvoire, where the public health clinics are nominally free (at least until the time of the survey), there is evidence that households living a greater distance from a clinic experience higher mortality among their children. Perhaps because of the more uniform distribution of child health benefits from the pub- lic health system in C6te d'Ivoire than in Ghana, household assets are not a significant predictor of child mortality in C6te d'lvoire. Advancements in women's education at the most basic levels in both countries are likely to foster further reductions In child mortality. To assess how maternal education at the three schooling levels is related to child mortality in the absence of other control vari- ables, a set of regressions was calculated with only the age of the mother con- trolled by the dummy variables (not reported here). Child mortality is signifi- cantly lower for each year of completed primary education of mothers, -0.011 in C6te d'lvoire and -0.008 in Ghana. These child survival benefits of maternal education continued and increased in magnitude in middle school in Ghana, -0.0 14, but lost their significance and size in C6te d'lvoire. No significant child health differences were associated with secondary or higher education in either country. 1,52 1 HF WORILD BANK ECONOMIC. RFVIFW. V:l. 10, NO. I One objective of this study was to consider the policy connections between child mortality and fertility. The available information on the local health pro- grams and environment could explain relatively little of the variation in child mortality across mothers in C6te d'lvoire and Ghana, although these instru- ments are statistically significant as a group. The Hausman (1978) test could not reject the hypothesis that child mortality is exogenous. When child mortal- ity is treated as an exogenous variable in fertility model I in the aggregate or disaggregated regressions, we estimate that four to fifteen fewer child deaths are associated with a reduction of only one birth. We have no good explanation for the small size of this estimate of the fertility response to child mortality. The study also sought to evaluate the other determinants of fertility in these two countries, which have very high fertility rates and high child mortality. Women's education, particularly beyond the primary school level, is strongly related to declines in fertility in both countries, but the education of husbands is not associated with similar declines. However, other measures of wealth and socioeconomic status appear to have opposite effects on fertility in the two coun- tries. In C6te d'lvoire, assets and maternal health are positively related to fertil- ity, but in Ghana these variables are negatively related to fertility. An implica- tion of our model is that household income should be treated as an endogenous variable. This led us to omit income from this study. In another investigation of these data, IV estimates of income effects, identified by household assets, are substantial in both countries, but positive in Cote d'lvoire and negative in Ghana (Benefo and Schultz 1994). Age disaggregation of rural and urban subsamples suggests that the negative effect of household assets on fertility is only evident in the urban subsamples of both countries among older women. The variables examined here are undoubtedly measured with error, and the rela- tionships estimated omit many relevant factors. It is also risky to infer how time trends will evolve in a society even from well-measured cross-sectional patterns. Nonetheless, there are several similarities and differences between these countries that may help forecast future trends. In C6te d'lvoire, increments to household nonhuman wealth are associated with higher fertility, while in Ghana the tendency to invest family wealth in having more children has been altered, at least in urban areas. Although per capita incomes and male earnings were higher in Cote d'lvoire than in Ghana, the earlier investmenit of Ghania in an egalitarian educational system has provided womiien witlh greater productive opportunities relative to men than those of women in C6te d'lvoire. Given the relationship observed between female education, wages, and productivity and reduced fertility, we expect that the chang- ing composition of income sources in Ghania will be more favorable for women and hence contribute to an earlier national decline in fertility than in C6te d'lvoire. Urbanization in both countries is likely to lower national fertility levels, but with the greater integrationi of the regional labor markets in Ghana, the potential for urban- ization to foster a fertility decline is probably greater in Cote d'lvoire. A more equal distributioll of social services would appear likely to hasten the decline in child morttality and fertility, particularly if women's education in- Benefo and Schultz 15.3 creases more rapidly in rural areas and rural sanitation and health problems are effectively addressed. Only in the case of community access to protected water supplies did we find evidence that, without a prior investment in female educa- tion, improvements in water supplies are not associated with increased child survival among the rural poor. A resumption of growth in personal incomes in Cote d'lvoire may offset, rather than reinforce, the fertility-reducing effect of the slow expansion in women's education. Conversely, sustained income growth in Ghana may benefit women's productivity as much as it does men's, and both income growth and increased education of women will work together to reduce childbearing and to shift social resources toward greater investments in child quality in the form of schooling, health, and migration. Table A-1. Means and Standard Deviations of Variables, C6te d'lvoirc and Ghana Co3te d'lvoire Ghana Standard Standard Variable Mean deviation Mean deviation Dependent variables Number of children born alive 5.48 2.85 4.72 2.47 Nuinber of children born in last five sears 0.92(l 1.00 1.11 0.934 Proportion of children born at least five 0.161 0.233 0.157 0.254 years before the survey who died before their fifth birthday (child mortality rate) Infant death rate (before first birthday) 0.l 18 0.209 0.114 0.220 Explanatory variables at individual household level Woman's schooling (years) Primary (1-6) 1.0( 2.15 2.76 2.86 Middle (1-4) 0.232 0.870 1.25 1.76 Seconidary and higher education 0.0602 0.619 0.233 1.29 (1-3 or more) Woman's age (years) 39.6 13.5 34.0 7.74 Woman's heighr (In meters) 0.45 0.0390 0.455 0.0393 Household assets (local currencv) 16.5 176 70.6 339 Share of tree crops" 0.346 0.298 0.295 0.259 Woman migrant (dummy)' 0.357 0.441 No husband present (dummy) 0.24( 0.354 Woman head of household (dummy) (1.080 0.290 Houisebold composition and husband characteristics5 Husband's schooling (years) Primary (1-6) 1.31 2.40 2.30 2.87 Middle (1-4) 0.4(06 1.16 1.29 1.83 Secondary aid higher education 0.199 1.04 0.424 1.71 (1-3 or more) Husband's height (In meters) 0.393 0.224 0.338 0.253 Husband's age (years) 36.9 23.6 27.1 21.9 (Table continues o,i tbe fOllou.ing page.) 154 tHE WORLD BANK ECONOMIC RFVIFW, VOL. 10, NO. I Table A-1. (continued) C6te d'Ivoire Ghana Standard Standard Variable Mean deviation Mean deviation Other variables averaged for sample cluster Rainfall (centimeters per year)e 107.0 18.4 50.7 14.3 Distance to nearest marketplace 2.37 4.64 3.15 6.66 (kilometers for C/ite d'lvoire, miles for Ghana) Proportion of cluster sample households 0.580 0.400 0.569 0.358 with toilet or latrine Proportion of cluster sample houselholds 0.491 0.352 0.337 0.395 with protected water source including piped water and wells with pumps One of the two most serious community health problems (dummy variables) Malaria 0.103 0.443 Diarrhea 0.198 0.168 Measles, chickenpox, or other 0.176 0.270 infectious illnesses Distance to nearest health clinic 11.9 1 8.2 4.22 6.82 (kilometers for C6te d'lvoire, miles for Ghana) Child immunization campaign in last - 0.502 five years (dummy) Public health expenditures per person - 309.0 95.0 in the province (x101, 1988 cedis) Prices in the comnun,iWty- Manioc/cassava 0.0736 0.0458 26.6 4.94 Maize - 61.8 8.68 Fish/fish 0.437 0.169 526.0 118.0 Beef/eggs 0.81() 0.146 24.5 3.43 Palm oil 0.682 0.319 - Peanut butter 0.396 0.156 - Sugar - 152.0 19.1 Bananas 0.082(0 0.0410 - Antibiotics - 4.17 0.548 Woman 's religion (dtomm) variables) Muslim - 0.135 Christi an - 0.633 Traditional religions - 0.179 Woman's ethnicitv or languiage (dumnmy variables) Akan 0.291 0.464 Ewe n.a. 0.173 Ga-Adangbe n.a. 0.080 Dagbani n.a. 0.037 Hausa n.a. 0.024 Nzema n.a. 0.009 Other languages n.a. 0.213 Krou 0.089 n.a. Benefo and Schultz I55 Table A-1. (continued) Cdte dilvoire Ghana Standard Standard Variable Mean deviation Mean deviation Mande-North 0.090 n.a. Mande-South 0.137 n.a. Voltaic 0.097 n.a. Alien 0.13S n.a. Woman's cuirrent residenee (dummy variahles)g h Current rural resident 0.560 0.398 Abidjan/Accra 0.187 0.129 Other urban/urban coast 0.214 0.146 Urban forest n.a. 0.257 Urban savannalh n .. 0.070 East forest/rural coast 0.244 0.081 West forest/rural forest 0.142 0.173 Rural savaninah 0.213 0.144 - Not available. n.a. Not applicable. Note: Samples are reduced by about 5 percent to hase complete reporting of height, and by 30 percent to include only women with at least one birth five or more years ago. The reproductive module was adniinistered to one woman in each household who was between age fifteen and fifty in Ghana and fifteen or older in CGte d'Ivoire. Thus, the sample from Cote d'Ivoire is older than that in Ghana by almost five years. Standard deviations are not reported for dummy variables because they are equal to [mi(m - 1 )l'], where m is the mean. The sample sizes are 1,943 for Cote d'lvoire and 2,237 for Ghana. a. The value of the household's assets includes the valuie of owned land, business assets, and ten times other property incomile per year per adult, in thousanlds of (VA franics for Cote d'lvoire anid in thousands of cedis for Ghana. h. Proportioni of Ciluster sample household land area that is farmed in tree crops such as cocoa, coffee, bananias, and cocon.ts. c. A woman migrant is a rural-born, currentlv urban resident who has lived in an urban area for more than five years. d. The sample average is based on a variable that is set to zero for women svho have no reporting husband. For example, the average number of years of hushand's primary education for women with husbanids in CCOte d'lvoire is (1.31)/1 -0.240) = 1.72 years. e. Rainfall in the area, or at the nearest weather station, in annual average centimeters per year in Ghana and in centimeters for the presious year in Cote d'lvoire. f. Local prices averaged and adjusted for inflationi to the same date in local currency. g. The term befiire the slash (/) applies to Cite d'lvoire; the term after applies to Ghana. h. The excluded term is Abidjan or Accra. Source: Authors' ca[lcLlarions. 156 THE WORLD (WANK lCONONIK RrVIVEW, VOL.. 10, NO. I REFERENCES The word "processed" describes informally reproduced works that may not be com- monly available through library sysrems. Ainsworth, Martha. 1989. Socioeconomzic Determzinants of Fertility in C6te d'lvoire. LSMS Working Paper 53. Washington, D.C.: World Bank. Ainsworth, Marthla, and Juan MuLnoz. 1986. The C6te d'lvoire Living Standards Sur- uey. LSMS Working Paper 26. Washington, D.C.: World Bank. Ainsworth, Martha, Kathleeii Beegle, and Andrew Nyamete. 1996. "The Impact of Woomen's Schooling on Fertility and Contraceptive Use: A Study of Fourteen Sub- Saharan African Countries." Tl7e World Banlk Economic Review 10(1):85-122. Aly, Hassan Y., and Richard Grabowski. 1990. "Education and Child Mortality in Egypt." World D)evelopment 18(5):733-42. Barbieri, Magali. 1989. "The Determinants of Infant and Child Mortality in Senegal: An Analysis of nmis D)ata." Ph.D. diss., University of California at Berkeley, Depart- ment of Demograiphy, Berkeley. Barrera, Albino. 19990. "The Role of Maternal Schooling and Its Interactioni with Public Health Programs in Child Health Production." Journal of Development Economics 32:69-91. . 1991. "The Interactive Effects of Mother's Schooling and Unsupplemented Breastfeeding on Child Health." Journal of Development Economnics 34:81-98. Benefo, K. D., and T. P. Schultz. 1994. Determinants of Fertility and Child Mortality in C(.te ivoire. ISNS Working Paper 103. Washingtonn, D.C.: World Bank. Ben Porath, Yoruni. 1978. "Fertility Response to Child Mortality: Microdata from Israel." In S. H. Preston, ed., The Effects of Intlat and Child Mortality on Fertility. New York: Academic Press. Bound, John, D. A. laeger, and R. M. Baker. 1995. "Problems with Instrumental Vari- ables Estimation When the Correlation between the Instruments and Endogenous Explanatory Variables Is Weak." Journal of the American Statistical Association 90(430):443-50. Caldwell, J. C., and Pat Caldwell. 1987. "The Cultural Context of High Fertility in Sub-Saharan Africa." Population and Development Reviewt/ 13(3):409-37. Cantrelle, P., B. Ferry, and J. Mondot. 1978. "Relationship between Fertility and Mor- tality in Tropical Africa." In S. H. Preston, ed., The Effects of Infant and Child Mortality on Fertility. New York: Academic Press. Cochranie, S. H., andl K. C. Zachariah. 1983. Infaint and Child Mortality as a Determi- nant of Fertility. WorILd Bank Staff Working Paper 556. Washington, D.C.: World Bank. Cochranie, S. H., D. J. O'Hara, and Joanie Leslie. 197(0. The Effec-ts of Education on Health. World Bank Staff Working Paper 405. Washington, D.C.: World Bank. Feachemii, R. G., and D. 1T. Jamison. 1991. Disease and Mortality in Sub-Sabaran Af- rica. New York: Oxford Ulniiversity Press. Fogel. R. W. 1991. "New SoLirces and New Techniques for the Study of Secular Trends in Nutritional Status, Health Mortality, and the Process of Aging." NBER Working Paper 26. Nationial Bureau of Economic Research, Cambridge, Mass. Processed. Benefo and Schultz 157 Frank, Odile, and M. Dakuyo. 1985. "Child Survival in Sub-Saharan Africa: Structural Means and Individual Capacity." Center for Policy Studies Working Paper 122. The Population Council, Center for Policy Studies, New York. Processed. Freedman, Ronald. 1975. The Sociology of Human Fertility. New York: John Wiley and Sons. Glewwe, Paul. 1987. The Distribution of Welfare in the Republic of Cote d'Ivoire. L-sms Working Paper 29. Washington, D.C.: World Bank. Hausman, Jerry. 1978. "Specification Tests in Econometrics." Econometrica 47(1): 153- 62. Kritz, M. M., and D. T. Gurak, 1989. "Women's Statis, Education and Family Formation in Sub-Saharan Africa." International Fancily Planning Perspectives 15(3):100-05. Lee, B. S., and 'F. P. Schultz. 1982. "Implications of Child Mortality Reductions for Fertility and Population Growth in Korea." journal of Economic Development 7(1):21-44. Maglad, N. E. 1994. "Fertility in Rural Sudan: The Effect of lIandholding and Child Mortality." Economic Development and Cultural Change 42(4):761-72. Morrow, R. H., P. G. Smith, and K. P. Nimo. 1982. "Assessing the Impact of [isease." World Health Forum 3(3):331-35. Notestein, Frank. 1945. "Population-The Long View." In T. W. Schultz, ed., Food for the World. Chicago: Ulniversity of Chicago Press. Okojie, C. E. E. 1991. "Fertility Response to Child Survival in Nigeria: An Analysis of Microdata from Bendel State." In T. P. Schultz, ed.. Research in Population Eco- nomics. Vol. 7. Greenwich, Conn.: JA] Press. Oliver, Raylynn. 1995. Contraceptive Use in Ghana: The Role of Service Availability, Quality, and Price. I.sMS Working Paper Ill. Washington, D.C.: World Bank. Olsen, R. J. 1980. "Estimating the Effects of Clhild Mortality on the Number of Births." Demographly 17(4):429-44. Patterson, K. D. 1981. Heealth in Colonial Ghana: Disease, Medicine, and Socio- Economic Change, 1900-1955. Waltham. Mass.: Cross-Roads Press. Pitt, Mark. 1995. WRomien's Schooling, Selective Fertilitv atnd Child Mortality in Sub- Saharani Africa. l Ms Working Paper 119. Washington, D.C.: World Bank. Rosenzweig, M. R., and T. P. Schultz. 1982. "Determinants of Fertility and Child Mor- tality in Colombia." Report to USAID on grant L)SPE-(;-0013. Yale UIniversitv, Eco- nomic Growth Center, New Haven, Conn. Processed. - 1983. "Consumer Demand and Household Production: The Relationship be- tween Fertility and C;hild Mortality." American Econom-nic Review 73(2):38-42. Rosenzweig, M. R., and Kenneth Wolpin. 1986. "Evaluating the Effects of Optimally Distributed Programs." American Economic Revietw, 76(3):470-82. Schultz, T. P. 1969. "An Economic M)lodel of Family Planning and Fertility." Journal of Political Economy 77(2):153-80. 1976. "Interrelationships between Mortality and Fertility." In R. G. Ridker, ed., Population and Development: The Search for Selective Interventions. Baltimore: Johns Hopkins tUniversity Press. 1980. "An Economic Interpretation of the Decline in Fertility in a Rapidly Developing Country." In R. A. Easterlin, ed., Population and Economic Change in Detveloping Counttries. Chicago: University of Chicago Press. 158 THE WORLD BANK ECONOMIC REVIEW, VOL. 10, NO. I . 1981. Economics of Population. Reading, Mass.: Addison-Wesley. . 1988a. "Economic Demography and Development: New Directions in an Old Field." In Gustav Ranis and T. P. Schultz, eds., The State of Development Econom- ics: Progress and Perspectives. London: Basil Blackwell. . 1988b. "Population Programs: Measuring Their Impact on Fertility." Journal of Policy Modeling 10(1):113-49. 1992. "Assessing Family Planning Cost-Effectiveness: Applicability of Individual Demand-Program Supply Framework." In J. R. Phillips and J. A. Ross, eds., Family Planning Programmzes and Fertility. Oxford, U.K.: Oxford University Press. . 1994. "Marital Status and Fertility in the United States." Journal of Human Resources 29(2):637-69. . 1995. Returns to Reproducible Human Capital and Development. New Haven, Conn.: Yale University, Economic Growth Center. Schultz, T. P., and Aysit Tansel. 1992. "Measurement of Returns to Adult Health: Morbidity Effects on Wage Rates in C6te d'Ivoire and Ghana." Economic Growth Center Discussion Paper 663. Yale University, New Haven, Conn. Processed. Smith, Adam. 1961. The Wealth of Nations. Edward Cannon, ed. London: University Paperbacks, Methuen. Taylor, C. E., J. S. Newman, and N. U. Kelly. 1976. "The Child Survival Hypothesis." Population Studies 30(2):263-78. van de Walle, Etienne, and Andrew Foster. 1990. Fertility Decline in Africa: Assess- ment and Prospects. World Bank Technical Paper 125. Washington, D.C. van der Gaag, Jacques, and Wim Vijverberg. 1987. Wage Determinants in C6ted'Ivoire. LSMS Working Paper 33. Washington, D.C.: World Bank. Willis, R. J. 1974. "Economic Theory of Fertility Behavior." In T. W. Schultz, ed., Econonmics of the Family. Chicago: University of Chicago Press. World Bank. 1986. World Development Report 1986. New York: Oxford University Press. Contraceptive Use and the Quality, Price, and Availability of Family Planning in Nigeria Bamikale J. Feyisetan and Martha Ainsworth Nigeria has experienced high fertility and rapid population growsth for at least the past thirty years. Only recently have public auithorities launched efforts to promzote contraceptive use. In this article, individual w omen are linked to the characteristics of the nearest health facility, pharmacy, and souirce of family planning to assess the relative importance of women's socioeconomic background and the characteristics of nearby services on contraceptive uise. The results suggest that the limited levels of female schooling (and probably other factors affecting iwomen's opportunity cost of time) are constraining contraceptive use, especially in rural areas. Another major constraint to increased contraceptive use is the low iaailahbility of family planning services in Nigeria. Broader availability of the pill and other methods in pharmacies and of injectables and intrauterine devices (oui)s) in health facilities is likely to raise contraceptive use. Outpatient or consultationi f1ees at nearby health facilities do not appear to be constraining denmand for niodern -onitraceptive metbods. With a total population of nearly 100 million inhabitants, Nigeria is home to about one in every five Sub-Saharan Africans (World Bank 1992). Althouglh there are indications of a fertility decline in southwest Nigeria (Caldwell, Orubuloye, and Caldwell 1992), the country as a whole has experienced high fertility and rapid population growth for at least the past thirty years. Only relatively recently have public authorities become interested in affecting these trends and promoted contraceptive use. Among those involved in the delivery of family planning services, there is a broadly held conviction that improved avail- ability of contraceptives and higher quality of services will result in greater con- traceptive use. At the same time, even the most recent surveys indicate that Nigerians often prefer large families. In the 1990 Nigeria Demographic and Health Survey (NDHS), for example, the mean desired family size among the 40 Bamikale .1 Fevisetan is with the Department of Demography and Social Statistics at Obafeemi Awolowo University, Ile-Ife, Nigeria; Martha Ainsworth is with the Policy Research Department at the World Bank. This article was written as backgrounid for the reseaircih project on "The Economic and Policy Determinants of Fertility in Sub-Saharan Africa," finalced by the World Bank Research Committee IRPO 67691) and sponsored by the Africa Technical Department and the Policy Research Department of the World Bank. The authors gratefully acknowledge the late Esther Boohene, Trevor Croft of the Demographic and Health Surveys, and Kathleen Beegle and Susmita Ghosh. The authors appreciate comimiiients from John Caldwell, Susani Cochrane, J. A. Ebigbola, Andrew Foster, Elizabetlh Frankenberg, S. K. Kwafo. Samson ILamlenn, Paulina Makinwa-Adebusoye, Lewvis NdhlovuL, David Radel, Fred Sai. and Baha Traore. D 1996 The International Bank for Reconstruction anid Development /[IF WORI ) B\AN I i9 160 THE WORLD BANK ECONOlMI( RFVIFW. YVOL. 0. NO. I percent of women who gave numerical answers was 5.8 children (Federal Office of Statistics and [RD/Macro International 1992). About 60 percent of the re- spondents replied that their family size was "up to God," indicating that they have no preferences, they are reluctant to state their preferences, or they would like as many children as possible. The success of efforts to lower fertility and promote greater contraceptive use will depend on an understanding of the importance of factors affecting the de- mand for children and the demand for contraception-including individual char- acteristics and the availability, price, and quality of services. Public policy po- tentially can influence outcomes through all of these channels. However, up to now, an assessment of the impact of the availability, price, and quality of ser- vices on the demand for contraception in Nigeria has been hampered by the unavailability of adequate data. Demographic surveys have concentrated on the collection of data on household and individual characteristics that are hypothe- sized to influence fertility or family planning decisions. Community data, where collected, are often not linked with individual data because they are usually collected by different agencies or researchers for different purposes. The Service Availability Module attached to the 1990 NDHS has now made such an analysis possible. In this article, we link individual women with the characteristics of the near- est pharmacy, health facility, and source of family planning to assess the relative importance of socioeconomic background and service characteristics on contra- ceptive use in Nigeria. Section I presents background information on Nigeria's fertility trends, population policy, and economy. Section II describes the model that motivates the choice of variables in the empirical analysis. Section III de- scribes the women and facilities in our sample. Empirical results for contracep- tive use are presented in section IV. Section V summarizes salient findings and offers tentative conclusions. I. BACKGROlIND Nigeria has been characterized by high, yet stable, birth rates. The 1965/66 National Rural I)emographic Sample Survey gave a crude birth rate of 50 per 1,000 persons and an average completed family size of 5.6 children (Federal Office of Statistics 1968). United Nations' estimates put the total fertility rate (TFR) at close to 7 and the crude birth rate at 50 between 1960 and 1980, and did not indicate significant fertility declines over that period (United Nations 1985).1 More recently, the 1981/82 Nigeria Fertility Survey (NFS) found a TFR for Nige- ria of 5.94 in 1980-82 and the 1990 NDHS put the TFR at 6.01 in the period 1988-90 (National Population Bureau (Nigeria) and World Fertility Survey 1984; Federal Office of Statistics and [RD/Macro International 1992). 1. These are estimates from the "medium variant" projection. Other scattered surveys, which vary in scoipe and content, also indicare average completed famoily size of hetween five and six and crude birth rates around fifty (tor more details, see Arowolo 1984). Fevisetan and Ainsworth 161 Regional variations in fertility, especially between the southern and northern regions, have become more pronounced since 1980. The NFS found total fertility to be highest in the northwest (6.38) and southwest (6.25), lower in the north- east (5.95), and lowest in the southeast (5.72). By 1988-90, the TFRS in the southeast and southwest (5.46 and 5.57, respectively) were both lower by about one child than those in the northeast and northwest (6.53 and 6.64, respec- tively). There is evidence, therefore, of significant decline in fertility in south- western Nigeria in the 1980s. As Caldwell, Orubuloye, and Caldwell (1992) point out, this decline is important on a continental scale because the southwest- ern region of Nigeria is more populous than all but three Sub-Saharan countries. Note also that the difference between the lowest and highest TFRs has increased from 0.66 to 1.18 during the 1980s. Fertility differentials were also found by levels of education, place of residence (urban or rural), marital status, employ- ment status, and use of contraception, among others. With respect to education, the NDHS found that total fertility was lower among women who had completed secondary schooling (4.2) than among those who had completed primary school- ing (5.6) or those with no schooling (6.5). Women with incomplete primary schooling had the highest total fertility (7.2). The relatively stable high birth rates in Nigeria have been accompanied by steady declines in death rates. The crude death rate was estimated to decline from 25 in 1960 to 17 in 1983; during the same period, the infant mortality rate declined from 195 to 115 and the under-five mortality rate from about 325 to 190 per 1,000 live births (UNIcEF 1985). The 1990 NDHS estimated an even lover infant mortality rate of 91; the highest rate was in the northwest (110), while infant mortality in the other three regions ranged from 83 to 88 (Federal Office of Statistics and ]RD/Macro International 1992). The balance between the stable high birth rates and the steadily declining death rates (in the absence of a major contribution by net migration) has led to an annual population growth rate in excess of 3 percent. Eco nornv Before the 1970s, Nigeria depended mainly on agriculture for its domestic and foreign earnings. Five major cash crops were exported-cocoa, rubber, cot- ton, groundnuts, and palm products. However, oil was discovered in the late 1960s and went into large-scale production in the early 1970s. It became the major revenue earner and contributed immensely to the growth of Nigeria's economy during the early 1970s, as the country benefited from the high price of oil in the world market. This brought about a rapid increase in the number of educational institutions and health facilities and in the provision of roads, elec- tricity, and piped water. In addition, wages increased in the nonagricultural sector. By the early 1980s, dwindling oil revenues provoked a downturn in the economy (National Research Council 1993). Living standards worsened, as goods became unaffordable. The quality of health and education services deteriorated 162 THE WORLD BANK ECONOMI( RFVIEW, VOI.l I. NO. I and essential facilities became scarce. Primary enrollment rates peaked at 98 percent in 1980 and declined thereafter to only 72 percent in 1990 (Scribner 1995). The situation eventually led to the adoption of a structural adjustment program in 1986. By that time it had become very apparent that unchecked rapid population growth is undesirable, no matter the level of economic growth. With the assistance of several international agencies, the federal government set in motion a course of action to slow down the rate of population growth. Population Policy During the oil boom of the 1970s, rapid population growth was not perceived as an obstacle to economic growth. The Third National Population Policy reads: "Although Nigeria has (by world standards) a large and rapidly growing popu- lation, these demographic factors do not appear as yet to constitute a significant or serious obstacle to economic progress. The country is fortunate in possessing a large land area endowed with natural resources, which if carefully exploited should provide a basis for building a viable economy which would ensure a steadily rising standard of living for the population within the foreseeable future especially during the current phase of the country's demographic transition which is characterized hy rapid growth. . . ' (Federal Republic of Nigeria 1975: 293- 94). There was, however, a plan to contilnue with the integration of the family planning information and services into an overall health and social welfare sys- tem for the country through the National Population Council of Nigeria. As living standards worsened in the 1980s, official population policy changed. The period 1983-89 marked the beginning of a government-sponsored, national family planning program. The 1988 National Policy for Development, Unity, Progress and Self-Reliance acknowledged that the laissez-faire approach to popu- lation issues was not effective in lowering population growth and had adverse consequences on the welfare of the citizens and the socioeconomic development of the country. The new policy adopted specific demographic objectives and advocated extending coverage of family planning services to half of all women of childbearing age by 1995, and to 80 percent by 2(000 (Federal Republic of Nigeria 1988). Since 1983, organized family planning services have received great encour- agement from two related developments. First, the recognition of family plan- ning as part of the state public health system led to the establishment, in 1987, of a family planning coordinator in each state. Second, a major effort by the Federal Ministry of Health with the technical and financial assistance of the United States Agency for International Development (USAID), the World Bank, and the United Nations Population Fund (UNITPA) resulted in the distribution of contraceptives such as pills, injectables, iuDs, vaginal foaming tablets, and condoms to health facilities. Hospitals, health clinics, maternity centers or ma- ternity homes, family planning clinics, pharmacies, and patent medicine stores became involved in the distribution of these commodities. In addition to the organized stationary delivery points, the Planned Parenthood Federation of 1beyisetan and Ainswvorth 163 Nigeria has, in recent years, recruited individuals to distribute nonclinical contraceptives in rural communities. By 1990, 7.5 percent of all women of reproductive age were current users of contraception and use of modern methods was only 3.8 percent. Modern meth- ods include sterilization, IUD, injectables, pills, condoms, spermicides, and dia- phragm. The rate of use of any contraceptive among married women (6 percent) was only half that of unmarried women (13 percent) (Federal Office of Statistics and IRD/Macro International 1992). The probability of use of modern contra- ception rose with female education, from only 1.3 percent of women with no schooling, to 3.9 and 6.4 percent of women with incomplete and complete pri- mary schooling, respectively, to 16.7 percent of women who have completed secondary schooling. Caldwell, Orubuloye, and Caldwell (1992) point out that contraceptive use has particularly expanded among younger single women: the NDHS results found that 38 percent of use was among single women, and in their 1990 study of urban women in Ekiti (Ondo State), contraceptive use rates were three to five times higher among unmarried than among married women aged fifteen to twenty-four. The other main demand for modern contraceptives comes from married women who want to replace traditional methods of birth spacing. 11. THE MODEI. The demand for family planning is conditional on the demand for children. The "demand for children" broadly includes decisions on the number of chil- dren desired and the timing of births.2 Economic models of fertility highlight several key factors affecting the demand for children (Becker 1960, 1981; Rosenzweig and Schultz 1985): * The value of women's time, because childrearing is intensive in the use of women's time. * The price of other child inputs, such as food, clothing, schooling, and health care. * The value of children's time in home production now and as an economic asset in the future. * Household income. These are in addition to more subjective factors, often referred to as inherent "tastes" for children. The implication of these models is that as the costs of children rise, holding other factors constant, the number of children desired will decline. Alternatively, as the economic benefits of children rise, holding costs constant, fertility can be expected to rise. The effect of household (nonlabor) income is ambiguous, because higher income means that couples can afford more children or can decide to invest more per child for any given number. l. For a more formal exposition of the model, see ainnex I of Fevisetan and Ainsworth (1994h. 164 THE WORLD BANK E(CONOMIC REVIEW. VOL. 10, NO. I The demand for family planning services arises from their role as an input into producing and achieving a desired number of children. Therefore, family planning demand will depend on the price and quality of services, conditioned on the target number of children. The cost of family planning services includes the cost of the contraceptives, the cost of transport to the facility providing family planning services, the value of the time it takes to reach the facility, and the registration (or outpatient) fee. The quality of services can be proxied by such variables as the number and composition of staff, the number and types of services available, and the regularity of supplies of contraceptives. Other ways of preventing births-such as periodic abstinence, rhythm, and withdrawal- rely to an even greater extent than modern methods on behavior modification, and thus also entail costs to the user. The costs of different methods will deter- mine users' choices. When the cost of modern methods rises, we expect a shift from modern to traditional methods or, if traditional methods are also too costly to practice, to no contraception and possibly a higher family size. The empirical model estimated in this article expresses the use of modern contraception as a function of factors affecting both the demand for children and the costs of fertility regulation: (1) FP = FP(Pf,, Q, W, PC, I) where FP is use of modern contraception, PfP is the price of family planning services, Q is the quality of services, w is the opportunity cost of women's time, P1 is the price of other child inputs, and I is household nonlabor income. In the empirical work, we have employed variables that proxy the arguments in equation 1. We expect variables that measure the value of women's time, such as education, urban residence, and region of residence, to lead to lower demand for children and higher contraceptive use. Variables that reflect the price of family planning, such as the distance to a family planning source, the price of commodities, and registration, consultation, or outpatient fees, should lower the use of contraception. We expect that indicators of the quality of fam- ily planning services will raise contraceptive use. Examples of variables that might be perceived as reflecting quality include the type of family planning source (pharmacy, hospital, health clinic or maternity center, and health center), the number and composition of staff at the family planning source, the types of family planning services available, and the number and types of family planning methods available. We have no a priori expectations about the relation between contraceptive use and nonlabor income because it depends largely on the rela- tion between income and the demand for children. If, as incomes rise, women want more children, then we would expect a negative relation between income and the use of contraception. In economic terms, this is the case if children are "normal" goods. By contrast if, as incomes rise, women want fewer children (perhaps of higlher "quality"), then we would expect contraceptive use to rise with rising income. Feyisetan and Ainsworth 165 The left-hand side of equation I is measured by current use of a modern contraceptive. The dependent variable is dichotomous, taking a value of one or zero depending on whether the woman possesses the attribute under consider- ation. Estimating the coefficients of equation 1 using ordinary least squares (OLS) will produce biased estimates. Thus, maximum likelihood logit has been used to estimate the parameters of the equation. In order to interpret results of the relation between service characteristics and contraceptive use as an "impact" of services, we must assume that these services are randomly placed. However, if individuals selectively migrate to areas in re- sponse to program characteristics or if the placemiient of services is based on the characteristics of the population, then program placement would not be exog- enous and the estimates of the effect of program characteristics may be biased (Pitt, Rosenzweig, and Gibbons 1993; Rosenzweig and Wolpin 1988). For ex- ample, if public family planning services are selectively placed in areas with high fertility and low contraceptive use, one might observe a negative correlation be- tween the availability and use of family planning. On the other hand, if services are targeted to women in areas of high demand for contraception, then estimates of the positive relation between availability and contraceptive use may be over- stated. Unfortunately, with only a single cross-section of data to work with, we are unable to correct for the potential endogeneity of service placement. Family planning services are not widely available in Nigeria, however, so we suspect that services have not been targeted to areas with low demand. 111. DATA AND VARIABLE [)EFINITIONS This study uses data from the 1990 NDHS. Data were collected from 8,781 women on their socioeconomic characteristics, fertility histories, and use of con- traception, among others. (The sample design is described in detail in Federal Office of Statistics and IRD/Macro International 1992.) In addition, information on family planning services was collected from two sources: groups of four or five knowledgeable informants in the selected community, and staff at facilities visited by the interviewers. Informants for the community questionnaire were asked to identify the hospital, maternity clinic, health center, family planning clinic, and pharmacy nearest to each cluster of households interviewed for the NDHS. All named facilities (one of each type per cluster) were then visited for the service module if they were within six hours on foot from the cluster. At each facility, information was collected about the availability and costs of drugs and family planning methods, the types and number of health personnel, registra- tion or outpatient fees, and the types of maternal and child health services that were available. By linking the exogenous facility data to the individual data, we were afforded an opportunity to examine the impact of the availability, price, and quality of certain services on women's demand for modern contraception. The facility data have some limitations. First, the service availability sur- vey was conducted in only 1 85 of the 299 survev clusters, of which almost 166 TFIE WORLD BANK FCONOMIC REVIEAW, VOL. 10, NO. I all (165) were in rural areas. The criteria for including 20 urban clusters in the service survey were not documented, but roughly 10 to 15 percent of urban women from each of the four regions could be linked to the service survey. Because two-thirds of the women using contraceptives were living in clusters not surveyed by the facility module, we were not able to include them in our assessment of the impact of the quality of services on the de- mand for contraception. Second, the identification of health facilities and the estimation of distances were left entirely in the hands of the community informants. In situations where people want more from the government, there is the possibility of underestimating the number of health facilities in a local- ity. Respondents may also have identified facilities more distant than the nearest ones. Many clusters of women could not be linked to facilities be- cause the respondents could not name a facility, not because one was not available. Third, even when a facility was named by community respondents, it was not always interviewed. As a result, although 8,781 women were suc- cessfully interviewed, only 5,714 were located in clusters covered by the com- munity-facility survey and, of these, only 4,681 (81.9 percent) could be linked to the nearest health facility. Our main sample for analysis is the 4,589 women who could be linked to the nearest health facility and for whom all explana- tory variables at the individual and health-facility level were present. Characteristics of the Women Of the 4,589 women in the working sample, 32 percent reside in the north- east, 23 percent in the northwest, 29 percent in the southeast, and 16 percent in the southwest. Eighty-nine percent reside in rural areas. The average age of the women is 28.6 years; for ever-married women, the average age is 30.4 years. Approximately 84 percent have been married. Over half of the women (56 per- cent) are Muslim, 26 percent are Protestant, and 13 percent are Catholic. The average number of children ever born is 3.48. Approximately 23 percent of the women have no children, and 34 percent have 5 or more. The level of literacy of these mostly rural women is generally low. Only 38 percent have ever attended school, and the average woman in the sample has completed only 2.6 years of schooling. Among those who have ever attended school, the average years of schooling is 7. Among women who have husbands or partners, the male partners have on average 1 year more of schooling than the women (2.7 compared with 1.7 years). Slightly more than a third of the male partners had some schooling (35 percent) compared with only about a quarter of the women who had ever been married (28 percent). With respect to contraceptive use, 5.6 percent of the women in the working sample were currently using a traditional or modern method of contraception, and 2.9 percent were currently using a modern method. Of those currently using a modern method, 39 percent were using the pill, 22 percent IUD, and 14 percent injection (see table 1). The main sources of contraceptives were hospitals (41 percent) and pharmacies (25 percent). Fey:setan and Ainsworth 167 Table 1. Distribution of the Sample of Women Using Modern Contraception by Method and Source Method and source Percent Method Pill 38.9 [IUD 22.3 Injectioni 13.7 Condom 1 1.5 Sterilization 9.4 Spermicide, diaphragm 4.3 Total 100.0 Source Hospital 41.0 Pharmacy or patent medicine store 25.2 Health center, maternity center, or health clinic 13.7 Private clinic 7.2 Planned Parenthood clinic 2.9 Friends or relatives 2.9 Market 1.4 Private doctor 1.4 Husband's place of work 0.7 Unknown 3.6 Total 100.0 Note: The sample slLe is 139 woomen. Totals may not add to 100 hecause of rounding. Souirce: 1990 NDHS data. Health Facilities Community data were collected from 185 clusters of households (165 rural and 20 urban). The nearest of four types of stationary health facilities are iden- tified on the community questionnaire: hospital; health clinic, maternity center, or family planning clinic (including the few family planning clinics that were identified, this group is henceforth referred to as health clinic); health center; and pharmacy (see table 2). At least one type of health facility was identified in 184 clusters: all four different types were identified in 32 clusters, three types in 71 clusters, two types in 64 clusters, and one type in 17 clusters. The number of clusters identifying a facility is highest for the pharmacy (144), followed by the hospital (134), the health clinic (108), and the health center (102). The commu- nity informants reported that 52 percent of all the named facilities provide fam- ily planning services. The percentage providing family planning services varies among the different types of health facilities: hospital (81 percent), health clinic (55 percent), health center (45 percent), and pharmacy (28 percent). However, comparison of these responses with what was reported by the health facilities reveals that the community informants were often incorrect. Thus, in the analy- sis below we use measures of family planning availability collected directly from the facilities. The mean distance from the cluster to the nearest health facility varies by type of health facility and by availability of family planning services in the health 168 THF WORLD BANK ECONO1MI: REVIEW. \Oi. In. NO. I Table 2. Distance anid Time froni Communities to the Nearest Health Facilities Facility Health (aitlinic maternity center, or family Health Community-level variable Hospital planning clinic center Pharmacy Nearest facilitv Number of communities that namedl a facilitv 134 108 102 144 Percentage of all 185 clusters 72.4 .58.4 .55.1 77.8 Mean distance to the nearest facility (miles) 15.5 6.7 8.0 4.9 Mean travel time by the most common means of transport (minutes) 92.8 68.4 89.9 67.2 Nearest facility with family planning Number of facilities named with family planning 109 59 46 40 Percentage of all named faicilities 81.3 54.6 45.1 27.8 Mean distance ro the niearest named facilitv (miles) 14.9 7.6 7.6 5.4 Mean travel time by the most common meanls of transport (minutes) 92.8 87.9 84.7 235.2 Note: The community data are from 185 cluisters of households. Source: 1 990 NDIjo Service Availability Survey, communnity informants. facility. The mean distance is highest for the hospital (16 miles) and lowest for the pharmacy (5 miles). Hospitals and health centers with family planning ser- vices are slightly closer to the community than tihose without family planning, and health clinics and pharmacies with family planning clinics are slightly far- ther from the communities than those without family planning services. The average time it takes to reach the nearest facility by the commonest means of transportation is generally greater than one hour. It takes roughly an hour and a half to reach hospitals and health centers. This is not surprising because most of the sample is in rural areas. To be eligible for an interview, a facility had to be within six hours' walking distance from the community. A health facility of each type was visited if it met the eligibility requirement. Ninety-three hospitals, 91 health clinics, 88 health centers, and 127 pharmacies were thus visited (see table 3). In 165 clusters (ap- proximately 90 percent of the 185 clusters with community modules) at least I health facility was visited. Ninety-three percent of hospitals, 58 percent of health clinics, 61 percent of health centers, and 32 percent of pharmacies that were visited had family planning services. The private sector controls between 3 per- cent (health centers) and 40 percent (health clinics) of the facilities. The hospi- tals have been in existence for a longer period than the other types of facilities, and although higher proportions of hospitals charge a registration or outpatient fee, the fees are highest in the health clinics. Fevisetan and Ainswortl 169 Variations in the quantity and quality of medical personnel and inpatient services across types of health facilities reflect the level of complexity of the tasks each type of facility performs. Hospitals generally perform the most com- plex tasks, followed by health clinics. Although almost all of the hospitals have Table 3. Description of the Nearest Health Facilities Facility Health clinic inateruity center, or famtly Health Facility-level variable Hospital planning clinic center Pharmacy Number of facilities visited 93 91 88 127 Ownership (percent) Public 75.3 58.7 94.3 - Private 18.3 40.2 3.4 Other 6.5 1.1 2.3 - Mean years in operation 22.8 13.3 12.0 5.7 Prices Percentage of facilities that charge an outpatient fee 87.1 82.4 70.5 Mean registration fee (naira) among facilities that chargc' 2.55 4.00 1.71 Medical staff Percent with at least one doctor 95.7 52.2 27.3 - Mean number of doctors 11.6 1.1 0.5 - Mean number of nurses 40.6 3.3 3.4 - Mean number of midwives 27.0 3.3 3.7 - Infrastructure Mean number of beds 147.8 9.5 11.2 - Equipment (percent of facilities) Examination table 96.8 79.4 65.9 - Electricity 87.1 64.1 53.4 52.8 Refrigerator 92.5 44.0 53.4 26.0 Running water 8x5.0 53 .3 52.3 38.6 Services (percent of lacilities) Family planning 92.5 58.2 61.4 32.3 Postnatal care 95.7 89.0 72.7 - Antenatal care 94.6 75.8 63.8 Delivery services 95.7 86.8 60.2 - Antimalarial drugs 89.3 76.9 64.8 96.9 Antibiotics 76.3 51.7 47.7 60.6 Any vaccine 81.7 50.6 73.9 - - Not available. This information was not included in the questionnaire for pharmacies. a. The average 1990 market exchange rate was 12.8 naira per U.S. dollar. Source: 1990 NDH-S Service Availability Survey, facility respondents. 170 TIJE WORLD BANK ECONOMI( RkVIFNW, VOL.. 1D, No. I at least one doctor, only 52 percent of health clinics and 27 percent of health centers have one. On average, each hospital has about twelve doctors, health clinics one, and health centers zero. Similarly, the average number of nurses varies from three (health clinics and health centers) to forty-one (hospitals) and the number of midwives from three (health centers) to twenty-seven (hospitals) (table 3). With respect to infrastructure, the results show that examination tables, elec- tricity, running water, and refrigerators are more likely to be found in hospitals than in any other type of facility. The number of beds in each facility reflects the amount of inpatient service it provides. The hospitals have, on average, 148 beds; the average numbers for the health clinics and the health centers are about 10 and II, respectively. Information was also collected on the types of maternal and child health ser- vices that were available in each facility. Four types of services are identified here: antenatal, delivery, postnatal, and immunization. Hospitals are more likely to provide these services than any other type of facility. With respect to the supply of drugs, the pharmacies and the hospitals are the most likely to supply antimalarial drugs and antibiotics, respectively. It is not surprising that almost all the pharmacies had antimalarial drugs at the time of the survey. In Nigeria the patient (or the patient's relative) finds it miore convenient to visit a phar- macy for antimalarial drugs than to visit a hospital; hospitals are usually con- sulted for complicated cases. The Health Facilities with Family, Planning Services Table 4 describes health facilities that offer family planning services with respect to the family planning methods offered, the average price of such com- modities, and the regularity of their supply. The number of facilities with at least one doctor or nurse capable of inserting the ItlD exceeded the number with an IUD insertion kit at the time of the survey. This suggests that nonavailability of trained personnel is not the binding constraint for expanding availability of IUDs. The results also show that the likelihood of having sterilization performed is highest in the hospitals: 58 percent of hospitals with family planning services have at least one doctor trained in female sterilization. This figure contrasts with 32 percent for health clinics and 7 percent for health centers. With respect to the availability of methods, the pill is the most available, followed by the condom; female sterilization is least available. No method was available in all the facilities at the time of the survey. While the probability of obtaining the pill is highest in the health clinics, the probability of obtaining an IUD, condoms, or spermicides is highest in the hospitals. The health center is the second-best place where all methods, except female sterilization, can be found. The price of obtaining a method is generally higher in the health clinics and pharmacies, which are preponderantly privately owned. Not all facilities that had a method at the time of the survey always had the method. For example, although 87 percent of hospitals with family planning services had the pill at the Feyisetin and Ainsworth 1J71 Table 4. Description of the Nearest Health Facilities with Family Planning Facility Health clinic, maternity center, or familv Health Facility-level variable Hospital planning clinic center Pharmacy Number of facilities visited that offered family planning 86 53 54 41 Family planning methods offered (percent of facilities) Pills 87.2 94.3 92.6 68.3 Injection (e.g., Depo Provera) 64.0 7.7 68.5 .31.7 ILI) 86.1 64.2 68.5 - Condom 86.1 73.6 85.2 82.9 Foaming tablets (spermicide) 68.6 50.9 63.0 41.5 Female sterilization 46.5 17.0 3.7 - Mean price of mietbods (naira)' Pill 1.25 2.94 0.77 6.25 Injection (e.g., Depo Provera) 6.66 11.41 3.12 13.19 IUD 5.13 27.94 4.06 - Condom (.75 1.59 0.53 1.85 Foaming tablets (spermicide) 2.68 4.02 5.84 6.91 Female sterilization 25.94 86.22 3.70 - Percent of facilities that ran out of stock wvithin the six mnonths preceding the survev Pill 38.7 16.6 36.4) 3.7 Injeccion (e.g., Depo Provera) 21.8 22.2 37.8 46.2 IUD 17.6 3.1 18.9 - Condonis 16.2 16.2 23.9 23.5 Foaming tablets (spernmicide) 11.9 16.0 11.8 11.8 Percent with someone trained to insert iuD) 88.4 774 75.9 - Percent with an IDI) insertion kit in stock 86.1 '5.5 57.4 Percent with a doctor trained to perform sterilizations 58.1 32.1 7.4 -Not available. This information was riot included in the questionnaire tor pharmacies. a. The average 1990 market exchange rate was 12.8 naira per U.S. dollar. Mean prices inclide zeros. Source: 1 990 NDHs Service Availability SLirvey, facility respondenrs. time of the survey, only 61 percent had the pill in stock during the six months preceding the survey. Definitions of Variables The dependent variable is current use of any modern method of contracep- tion.3 The explanatory variables include characteristics of the women and char- 3. For results for determinants of ever use of contraception, see Fevisetan and Ainsworth (1994). 172 THE W()RLD BANK VCONOMI(. RlEVIEW. V\(L. 10, N(V I acteristics of the pharmacies, health facilities, and sources of family planning that are nearest to the women surveyed. The explanatory variables can be orga- nized into three groups: the individual woman's characteristics, variables re- flecting access to a health facility or the price of services, and those reflecting the quality of the nearest health facility. The individual woman's characteristics include age, education, place of residence, and religion. The choice of individual variables is based on their theoretical as well as practical relevance. Women's education is also inter- acted with urban residence and region of the country to see in which areas education has a greater effect. In the absence of a direct measure of wealth, the type of floor in the dwelling is adopted as an index on the assumption that women who live in dwellings with polished wood, vinyl, ceramic, or cement floors are likely to have higher income than those in dwellings with animal dung or earth-sand floors. Access or price variables include the distance to the nearest health facility, the availability of family planning services, and, conditional on the facility of- fering familv planning, the availability and price of individual methods. The quality of the nearest health facility is measured by whether it is privately owned, whether the facility has at least one doctor, and the number of contraceptive methods offered. IV. EMPIRICAL RESULTS The determinants of contraceptive use are estimated as a function of three types of facilities-the nearest health facility, the nearest pharmacy, and the nearest health facility offering family planning. Results are presented in tables 5, 6, and 7. Descriptive statistics for the variables in the regressions are in ap- pendix table A-i. Logit estimation results for the determinants of current contraceptive use as a function of the characteristics of the nearest health facility or nearest facility with family planning are presented in table 5. The nearest facility is defined as the hospital, health clinic-maternity center, or health center that is closest to a community. The first three specifications are for the sample of 4,589 women who could be linked to the nearest health facility and for whom there were no missing values on independent variables. The first specification shows the re- suilts for individual characteristics and characteristics of the nearest health facil- ity that reflect the quality, availability, and price of contraceptive methods. In the second specification, we add the characteristics of the nearest pharmacy. In the third specification, we replace the characteristics of the nearest health facil- ity by the characteristics of the nearest source of familv planning. The fourth specification is the same as the third, but includes the entire national sample of 8,761 women.4 Note, however, that although using the full sample adds obser- 4. The roral sample was 8,,7 1 women, hut values on the independenlt variables were missing for 20 women. Feyisetan and Ainsworth 1 73 vations to the variables measured at the individual level, the number of observa- tions on service characteristics is unchanged from the third specification and still reflects only the primarily rural areas where this information was collected. Thus, the fourth specification is shown primarily for purposes of comparing the results on women's characteristics with the earlier specifications. For ease of interpretationi, the logit coefficients have been transformed into the marginal effects of a change in the explanatory variables on the probability of contraceptive use, evaluated at the mean of each variable and multiplied by 100. Thus, the coefficients can be interpreted as the increase (or decrease) in current contraceptive use (in percentage points) associated with a one-unit change in the variable, evaluated at its mean value. Women's Charaicteristics In the first three specifications (table 5), current contraceptive use increases with age but at a decreasing rate. This is expected: as women age, the longer is their potential exposure to pregnancy and their fertility will approach desired family size. However, they also become less fecund and thus are less likely to need contraception to limit births. The two measures of the opportunity cost of women's time-education and urban residence-exert great positive impact on contraceptive use. Educationi and urban residence increase the opportunity cost of women's time by enabling women to hold higher-paying jobs, making childrearing more costly. This leads to a decline in the demand for children and an increase in the demand for contraception. The negative marginal effect of the interaction between schooling and urban residence indicates that the positive relation between female education and contraceptive use is weaker in urban than in rural areas. For ease of interpretationi of the relation between education and contraceptive use by region, table 6 combines the schooling coefficients interacted with regional and urban dummy variables using the first specification in table 5. The positive relation between female schooling and contraceptive use is greatest in the rural north and rural southwest. Although the interaction between female schooling and the regions is jointly significant, the dummy variables for regions are not individually or jointly sig- nificant in the first specification of table 5. However, when characteristics of the nearest pharmacy and nearest source of family planning are included (as in the second and third specifications), they become weakly jointly significant and women in the northwest appear to have lower contraceptive use than in other regions. In the first three specifications (table 5), women in Protestant house- holds are more likely to be current users of contraception. Most of the sample in the first three specifications is from rural Nigeria, where the most distinguishinig feature of wealth is the type of housing a family has. As such, women who reside in dwellings with polished wood, ceramic, or cement floors no doubt have access to higher income. The positive marginal effect on type of floor, therefore, indicates an increase in the demand for contra- ception as income rises. This result might seem to imply that income would 174 IHE :WORLD BANK ECONOMIC REVIEW, VOL. 1D, NO. I Table 5. The Marginal Effect of the Characteristics of Women and the Nearest Health Facility, Pharmacy, or Family Planning Source on Contraceptive Use Specification Explanatory v'ariable 1 2 3 4 Individual woman Age (years) 0.277 0.231 0.242 0.509 (3.19) (3.45) (3.48) (5.84) Age squared -0.003 -0.003 -0.003 -0.006 (-2.21) (-2.41) (-2.46) (-4.42) Urban residence' 1.252 1.276 1.421 2.298 (2.14) (2.79) (3.12) (4.21) Years of schooling 0.1.32 0.127 0.138 0.346 (3.1 7) (4.00) (4.27) (5.47) Interactiom: urban x schooling -0.094 -0.093 -0.118 -0.133 (- 1.81) (-2.40) (-2.72) (-2.03) Northeast residence' -0.196 -0.120 -0.352 -0.971 (-0.41) (-0.36) (-1.06) (-1.44) Northwest residence' -0.780 -0.743 -0.999 -2.213 (-1.38) (-1.61) (-2.23) (-2.82) Southwest residence' -0.177 -0.117 -0.330 -0.721 (-0. 32) (-0.23) (-0.74) (-1.20) Interaction variables Northeast x schooling 0.104 0.060 0.060 0.019 (1.37) (0.93) (0.83) (0.15) Northwest x schooling 0.107 0.067 0.094 0.198 (1.61) (1.33) (1.83) (2.20) Southwest x schooling 0.088 0.(72 0.083 0.034 (1.46) (1.55) (1.69) (0.45) Type of floorh 0.754 (.SII1 0.618 0.498 (2.19) (1.9.5) (2.37) (2.04) Protestant' 0.585 0.519 0.546 0.375 (2.12) (2.45) (2.41) (1.29) Muslim l 0.154 0.152 0.147 -0.105 (0.39! (0.51 ) (0.48) (-0.30) Nearest health facility Distance (miles) -0.088 -0.074 -0.036 -0.081 (-2.74) (-2.65) (-1.65) (-1.42) Outpatient registration fec -0.038 -0.()17 0.026 0.041 (-1.17) (-0.62) (0.84) (0.54) Family planning offered, -1.132 -1.212 (-1.73) (-2.62) Injection offered, 0.998 0.896 0.967 1.464 (2.29) (3.111) (3.69) (2.68) Injection price' -0.027 -0.049 -0.063 -0.072 (-0.56) (-1.92) (-3.38) (-1.91) IUD offered, 0.830 0.414 (1.69) (1.57) IL!D price,d -1.41x104 0.015 (-0.(01) 1.84) Pill offered, 0.105 0.578 -0.408 -1.057 (0.17) (1.6.3) (-1.04) (-1.17) Feyisetan and Ainsworth 1 75 Table 5. (continued) Specification Explanatorv variable 1 2 3 4 Pill price- d 0.028 0.054 0.154 0.251 (0.47) (0.97) (3.24) (2.45) Condom offered' 0.045 -0.057 0.422 0.794 (0.12) (-0.22) (1.46) (1.14) Condom price' d 0.065 0.1(12 0.140 0.247 (0.85) (2.03) (3.36) (2.42) At least one doctor -0.645 -0.743 -0.621 -1.421 (-1.89) (-2.83) (-2.42) (-2.23) Hospital' 0.694 0.910 0.575 1.767 (1.92) (3.24) (2.29) (2.67) Health cliniic 0.222 0.124 0.014 0.497 (0.62) (0.45) (0.04) (0.70) Privately owned facility' 0.639 (0.516 0.233 0.168 (1.66) (1.64) (0.87) (0.25) Health facility with family planning -0.130 0.094 linked to the cluster, (-0.58) (0).17) Nearest pharniacy Pharmacy interviewed, -1.446 -1.325 -2.087 (-3.64) (-2.97) (-1.88) Distance (miles) 0.011 -0.026 -0.109 (0.37) (-0.78) (-1.16) Number of hours open per week 0.014 0.016 0.022 (3.62) (3.61) (1.88) Family planninig offered, 0.333 -0.184 -0.314 (0.66) (-0.46) (-0.30) Niumber of family planning 0.541 0.666 1.703 methods available (1.68) (2.81) (3.12) Injection offered, -0.319 -0.137 -0.785 (-0.81) (-0.44) (-0.90) Injection price' d -0.034 -0.024 -0.063 (-3.08) (-2.84) (-2.78) Pill offered, 1.041 1.588 3.325 (1.86) (2.83) (2.51) Pill price',d -0.254 -0.321 -0.698 (-2.05) (-5.09) (-3.66) Condom offered' -0.630 -0.530 -1.537 (-1.09) (-0.97) (-1.08) Condom price'd -0.382 -0.417 -1.074 (-1.65) (-1.81) (-1.77) Constant -10.037 -10.441 -10.944 -9.265 (-7.73) (-7.77) (-8.48) (-13.17) Likelihood -486.75 -472.56 -475.36 -1465.45 PseudoR2 0.1962 0.2196 0.2150 0.1821 Sample size 4,589 4,589 4,589 8,761 (Table continues on the following page.) 176 THE WORLD BANK FCONOM(N REVIFW, VOI. 1), NO. I Table 5. (continzued) Specification Explanatorv variable 1 2 3 4 Joint testsC Schooling: regions x schooling 35.94 41.54 54.43 73.34 [0.0001 [0.000] l0.(001 [0.000] Method availability 15.50 16.94 16.44 8.03 (health facilitv) 10.0091 [0.00] [0.003] [(.0901 Method availabilty (pharmacyi 29.53 38.55 19.06 [0.0(01 [0.0001 [0.002] Note: Values represent the marginal change in current contraceptive use (percent) for a one-unit clhanige in the explanatory variable. t-statistics are in parentheses and are froni the original logit regressions. Logit coefficients have been transformed into marginal changes using the formula dpIdx5 = f3k exp (x'fl)/ 11 + exp x [2, and the result has been multiplied by 100. Standard errors have been estimated using Huber's techniique, which is robtist to heteroskedasticirv and cluster effects. a. Dummy variable: the value is I if the condition is true: 0 otherwise. b. Dummny variable: the valuie is I if the floor is parquet, vinyl, ceramic, or cement; 0 otherwise. c. Fees and prices are in naira. The average 1990 market exchanige rate was 12.8 naira per U.S. dollar. d. The variable is availability of the method ( I or 0) tines the price. e. The value reporteLl is the X- statistic for the joitnr test; p-values are in brackets. Source: 1990 Nt1115 data. exert a negative effect on fertility. However, this seems not to be the case in Nigeria (see Ainsworth, Beegle, and Nyamete 1995). The results for individual variables in the fourth specification reveal that the relationship betweeni contraceptive use and female schooling, area, and region of residence is stronger in the sample of all Nigerian women than in the sample of women for which the characteristics of health facilities were available. The marginal effect of an additional year of female schooling, for example, increases from roughly 0.1 3 in the first three specifications to 0.35 in the fourth (table 5). The relation with urban residence also increases dramatically from about 1.3 in the first three specifications to 2.3 in the fourth. These "stronger" results arise from the fact that the marginal effects are computed at the mean of the indepen- dent variable, and the mean levels of schooling and urban residence are higher for the sample of all Nigerian women. For example, the mean years of female schooling for the first three regressions is 2.6 years, while for all women it is 3.7 years. Eleven percent of the women in the first three regressions lived in urban areas, compared with 40 percent of those Nigeria-wide (see table A-1). Nearest Health Facility and Nearest Facility wvith Family Planning With respect to the characteristics of the nearest health facility, both the first and second specifications in table 5 show that greater distance is associated with a lower probability of contraceptive use, irrespective of whether family plan- ning is actually offered at the facility. When in the third specification we con- sider the distance to the nearest health facility with family planning, the relation is also negative, but only half as strong and very weakly statistically significant. Feyisetan and Ainswortb 177 Table 6. The Marginal Effect of a One-Year Increase in Female Schoolinig on Current Contraceptive Use Region lUrban Rural Southeast 0.04 0.13 Southwest 0.13 0.22 Northeast 0.14 0.24 Northwest 0.15 0.24 Note: The data show the marginal change in current contraceptive use for a one-unit change in the explanatory variable. Logit coefficients have been transformed into marginal chaniges using the formula dpldx, = , exp (x'fl)/[I + exp (x'l)12. and the result has been multiplied hy 100. Source: 1990 NDIIS data. In none of the specifications is current use of contraception related to the level of outpatient or registration fees. Conditional on the availability of any method of family planning, women with access to injections and the IUD at the nearest health facility are more likely to be current users (the first specification).S The coefficients on availability of other specific methods are not individually significant, but are jointly significant in determining current contraceptive use. The results for prices of specific meth- ods are generally insignificant, but are sensitive to inclusion of the characteris- tics of the nearest pharmacy in the second and third specifications. In the latter case, strong positive price effects on contraceptive use appear for the IUD, pill, and condom, while the relation with injection prices is always negative (table 5). Controlling for the other characteristics of health facilities, women for whom the nearest health facility or source of family planning is a hospital are more likely to be current users of contraception, compared with those for whom the nearest facility is a health clinic or dispensary. This may have to do with the quality of the services provided at a hospital, which we have been unable to completely control for, or may reflect some characteristic of women who live near hospitals. Women for whom the nearest health facility is private are also more likely to be using modern contraception, although whether or not the nearest facility with family planning is private is not re- lated to contraceptive use. Curiously, the presence of a medical doctor at the nearest facility is associated with lower contraceptive use and almost exactly offsets the positive coefficient on hospital. A similar result has been found in Zimbabwe (see Thomas and Maluccio 1995). One plausible explanationi is that most Nigerian doctors are men; women may prefer to receive contra- ceptive services from female providers. In other specifications, the distance to a primary school and the number of years that the facility had been oper- ating were not significant determinants of current use. 5. When the avallability of various individual methods and their prices are included in the regression. the coefficient on tlie availability of any family planning is negative (see table 5). However, thlis coefficienit must be used in conibination with those on specific methods. Wheni only the availabilitv of familv planning is entered (without dummy variables for specific merhods anid their prices), the coefficient is positive, although of borderline significance (see Feyisetan and Ainsworth 1994, tahle 6A, specificationi 21. 178 THE WORLD BANK ECONOMI( RI VIEW, VOL. IUI. NO. I Nearest Pharmacy Because of the recencv in the integration of family planning programs into the primary health care scheme, the pharmacy has, for a long time, been the major distributor of nonsurgical contraceptives in Nigeria. It is not uncommon in some areas to find pharmacies stocked with contraceptives that are lacking in the hospitals, especially in government-owned hospitals. But because pharma- cies do not perform several of the functions of the three other health facilities (hospitals, health clinics, and health centers), we present the results for phar- macy characteristics in addition to whichever of the three other stationary fa- cilities is closest to the community. In the second and third specifications in table 5, we add to the regressions the characteristics of the nearest pharmacy. The nearest pharmacy was not inter- viewed for 27 percent of the women for whom the characteristics of the nearest health facility were obtained. A dummy variable for whether a pharmacy was interviewed has thus been introduced. The distance to the nearest health facility (and, marginally, the nearest source of family planning) remains a significant determinant of contraceptive use, but the distance to the pharmacy is not sig- nificant. However, the results for other characteristics of the nearest pharmacy are more highly and consistently significant than those for the nearest health facility or nearest facility with family planning. Increased hours of operation at the nearest pharmacy are associated with higher contraceptive use, as is the number of methods offered. Among specific methods, availability of the pill at the nearest pharmacy is associated with significantly higher contraceptive use, but the prices of injections, the pill, and condoms are associated with lower contraceptive use. Recall that only 3 percent of women overall were current users of a modern method. Only 22 percent of the women who could be linked to a pharmacy lived nearest to a pharmacy that supplies the pill. This suggests that increased availability of the pill at private pharmacies may have an impor- tant impact on raising contraceptive use in Nigeria. Inithe fourth column of table 5 the full set of methods and quality character- istics is entered for the nearest family planning source and pharmacy for all women in the N[)HS, irrespective of whether the women could be linked to these facilities. The major difference between the results of the third model and those of the fourth is that the levels of significance of some of the characteristics of the nearest family planniing source and pharmacy decrease in the fourth model, while the size of the marginal effects increases. Note that the pseudo R2 declines for the full sample of women as well; this is not surprising in light of the fact that 70 percent of them had no values for the characteristics of the nearest source of family planning anid 56 percent had no pharmacy characteristics. Regionial Differences A chi-squared test confirmed that there are structural differences in the determi- nants of contraceptive use by region. The sixth and seventh specifications in table 7 Fevisetan and Ainswortb 179 show the relationship between the characteristics of the nearest health facility and current use of contraception among women in the sample living in the north and south. Because the models were estimated separately for the two regions, the inter- actions between urban residence, women's education, and regions have been dropped. The results of the fifth specification, on the pooled women and without these re- gional and schooling interactions, are provided for comparison. A strong relation between urban residence and contraceptive use is observed among northern women, but not among women in the south. The marginal effect of a one-year increase in women's schooling is also much larger in the north (.402) than in the south (.016). In the north, women's schooling has a positive relation with contraceptive use in both urban and rural areas, but in the south it shows a net positive relation in rural areas only. The proxy for income is insignificant among women in both regions, but it is significant for the entire sample in the fifth specifi- cation. Protestant women in the north are more likely to use modern contraception than women from other religious groups; both the significance of the coefficient and the size of the marginal effect of religion are smallcr in the south. There are important differences among northern and southern women with respect to the relation between characteristics of services and contraceptive use as well. Among nortlhern women, the distance to the nearest health facility, the price of outpatient consultations, and the presence of a medical doctor are sig- nificantly associated with lower use of moderni contraceptives. In addition, con- traceptive use is higher among women for whom the nearest facility is a hospi- tal. None of the variables measuring the availability or price of specific methods are statistically significant. In contrast, among southern women contraceptive use does not significantly vary by distance, the presence of a doctor, or the type of facility, and higher outpatient fees are associated with higher contraceptive use. The results for dis- tance and presence of a doctor reinforce the suspicion that the negative relation between doctor and contraceptive use may be caused by the gender of the doc- tors, because the majority of northern women are JMuslim. The positive relation between outpatient fee and contraceptive use among southern women may be capturing some other aspect of service quality that could not be controlled for but that is associated with higher fees. Finally, the availability of the pill is very strongly correlated with higher contraceptive use among women in the south, and private ownership of the nearest health facility is associated with lower use. All of the contraceptive methods are jointly significant for the southern sample, but only at marginal levels for the nortlherni sample. The independent variables explain a much higher proportion of the variation in contraceptive use among southern women than among northern womeni (the pseudo R2 is 0.2718 and 0.1457 for southern and northerni women, respectively). Partner's Characteristics In the first seven specifications, we have included all women in the regres- sions, regardless of their marital status. There are several reasons for this. First, 180 THE WORLD BANK ECONOMIC REVIEW, VOI. 1O, NO. I Table 7. The Marginal Effect of the Characteristics of Women and the Nearest Health Services on Current Contraceptive Use, by Region and Marital Status Women Women who in the Womeni in haue a All women north the south partner Explanatory variable 5 6 7 8 Individual woman Age (years) 0.281 1.013 -1.03xl04 0.281 (3.16) (4.07) (-(.01) (3.00) Age squared -0.003 -0.013 1.04xl0A -0.003 (-2.18) (-3.38) (2.46) (-2.07) Urban residence' 1.057 2.853 0.033 0.503 (1.98) (2.27) (0.32) (1.29) Years of schooling 0.187 0.402 0.016 0.120 (5.43) (5.45) (2.24) (3.01) Interaction: urban x schooling -0.075 -0.119 -0.016 (-1.66) (-1.00) (-2.66) Northeast residence' 0.278 0.355 ((1.66) (0.83) Northwest residence' -0.214 -0.133 (-0.47) (-0.29) Southwest residence' 0.380 0.376 (1.03) (1.01) Type of floor" 0.745 1.126 0.059 0.748 (2.10) (1.23) (1.30) (2.03) Protestant' 0.607 1.294 0.115 0.585 (2.18) (1.82) (1.56) (2.061 Muslim, 0.095 1.000 0.049 0.086 (0.24) (0.90) (0.61) (0.22) Partner or husband., -0.707 (-1.58) Partner's years of schooling 0.074 (2.32) Nearest health facility Distance (miles) -0.102 -0.139 -0.008 -0.105 (-2.98) (-1.72) (-1.52) (-3.08) Outpatient registration fee, -0.036 -0.237 0.028 -0.043 (-1.01) (-2.30) (3.08) (-1.18) Family planning offered < -1.178 -2.550 -1.169 -1.058 (-1.74) (-1.07) (-11.95) (-1.55) Injection offereda 0.968 1.095 0.063 0.830 (2.24) (0.78) (0.94) (2.07) Injection price; d -0.021 0.017 0.015 -0.015 (-0.41) (0.11) (1.01) (-0.33) IL) offered, 0.880 2.190 0.011 0.816 (1.74) (1.48) (0.13) (1.69) IUD price" -8.22x10' -0.015 0.003 -0.004 Feyisetan and Ainsworth 181 WXomen Women who in the Women in have a All women north the south partner Explanatorv variable 5 6 7 8 (-0.04) (-0.27) (0.27) (-0.20) Pill offered, 0.238 0.717 1.239 0.302 (0.37) (0.29) (14.19) (0.47) Pill price,.d 0.01.5 0.141 -0.004 0.008 (0.25) (0.97) (-0.17) (0.15) Condom offered, 0.060 -0.197 -0.116 -0.007 (0.16) (-0.19) (-1.64) (-0.02) Condom priced 0.044 0.096 -0.018 0.041 (0.58) (0.44) (-0.83) (0.58) At least one doctor' -0.697 -2.107 -0.047 -0.676 (-1.97) (-2.04) (-0.61) (-1.97) Hospital' 0.111 2.704 0.013 0.772 (1.85) (2.19) (0.17) (2.05) Health clinic, 0.241 1.361 -0.050 0.249 (0.66) (1.41) (-0.98) (0.70) Privately owned faLility' 0.731 0.987 -0.407 0.753 (1.88) (1.02) (-1.78) (1.93) Constant -10.304 -11.839 -7.747 -9.726 (-7.99) (-7.57) (-3.57) (-7.23) Likelihood -488.58 -339.70 -134.28 -486.57 PseudoR` 0.1931 0.1457 0.2718 0.1965 Sample size 4,589 2,071 2,518 4,589 Joint tests, Regions 2.63 2.59 [0.4521 [0.459] Method availability (health facility) 16.63 9.26 269.87 13.53 [0.005] 10.0991 [0.000] [0.019] Woman's schooling. partner, 41.40 partner's schooling [0.000] Note: Values represent the marginal change in current contraceptive use (percent) for a one-unit change in the explanatory variable. t-statistics are in parentheses and are from the original logit regressions. Logit coefficients have been transformed into marginal changes using the formula dpldx, = Ps exp (x jf)/ [I + exp (x f/)]2, and the result has been multiplied by 100. Standard errors have been estimated using Huber's technique, which is robust to heteroskedasticity and cluster effects. a. Dummy variable: the value is I if the condition is true; 0 otherwise. b. Dummy variable: the value is I if the floor is parquet, vinyl, ceramic, or cement; 0 otherwise. c. Fees and prices are in naira. The average 1990 market exchange rate was 12.8 naira per U.S. dollar. d. The variable is availability of the method ( I or 0) times the price. e. The value reported is the X2 statistic for the joint test; p-values are in brackets. Source: 1990 NDOIS data. 182 TUF. WORID BANK lCONONIK. REVIFEN', VOL. 10, No. I as Caldwell, Orubulove, and Caldwell (1992) have pointed out, the demand for contraception appears to be higher among unmarried women than among mar- ried women. Dropping the unmarried women would therefore eliminate a ma- jor share of users. Second, the decision to marry is related to the decision to have children, and thus marital status is jointly endogenous with fertility and contra- ceptive use. Third, to the extent that more-educated women delay marriage and therefore have higher contraceptive use before marriage, excluding unmarried women will distort the relation between education and contraceptive use. How- ever, one advantage of conditioning on marital status is the ability to compare the relative effects of male and female schooling on women's contraceptive use. Thus, in the last column of table 7, we control for whether the woman has ever had a husband or partner and for the education of the current or most recent partner. Having a partner is associated with lower contraceptive use, but the relation is not statistically significant at conventional levels. When the partner's or husband's variables are included, the marginal effect of female schooling de- clines (from 0.19 in the fifth specification to 0.12 in the eighth specification), but is still significant and substantially greater than the marginal effect of male schooling (0.07) (see table 7). Husbanid's schooling is often included as a proxy for income, but its inclusion in the eighth specification does not detract from the proxy for wealth (type of flooring), which remains positive and significant. The results for other characteristics of health facilities are substantially the same as in the fifth specificationl. V. SUMIMIARY ANID CONCLUSIONS This article has examined the impact of individual characteristics and the availability, price, and quality of services at the nearest pharmacy and health facility on current use of contraception in Nigeria. Only 3 percent of the women in this study were currently using a modern method of contraception. Among the background characteristics that had the strongest relation with contracep- tive use are factors thought to raise the opportunity cost of women's time- women's schooling and urban residence. Women's schooling in particular had a consistently strong positive relation with contraceptive use in every specifica- tion; in rural areas, the relation between schooling and contraceptive use is more pronounced. In the sample of women analyzed, overall levels of schooling were extremely low. Of all women aged fifteen to forty-nine in our sample, 62 per- cent had received no formal schooling whatsoever; in rural areas, this figure is 67 percent, and in urbani areas about a quarter of the women had no schooling. Our analysis strongly suggests that policies to raise female schooling will result in greater contraceptive use, holding constant the characteristics of services. The relation betweeni our proxy for income and contraception was also consis- tently positive, iiidicating that higher incomes will contribute to greater contra- ceptive use. Fevisetan and Ainsi'orth 183 The low availability of family planning services in Nigeria is a constraint to increased contraceptive use. Increased distance to the nearest health facility was associated with lower contraceptive use, but distance to the nearest facility with family planning was not. This suggests that women who seek health care are more likely to consider using contraception, even if methods are not available at the nearest facility. Women in the sample were generally closer to a pharmacy (4.3 miles) than to the nearest health facility (5.1 miles) or to the nearest facility with family planning (6.9 miles), making the pharmacy the least costly source in at least one sense. However, pharmacies in this predominantly rural sample were also less likely to stock contraceptives than were health facilities: 32 per- cent of the nearest pharmacies offered family planning, but 55 percent of the nearest health facilities did so. Longer hours of operation at the nearest phar- macy were associated with higher use. Conditional on the availability of any family planning method at the nearest health facility, the availabilitv of injec- tions and sometimes the IUD is often significantly associated with higher current use of contraception. At pharmacies, the total number of methods and the avail- ability of the pill are associated with higher contraceptive use, and the price of all methods is associated with lower use. Outpatient or consultation fees at nearbv facilities do not appear to be constraining demand for modern methods. The results of this study suggest that the limited levels of female schooling (and probably other factors affecting women's opportunity costs of time) are constraining contraceptive use in both urban and rural areas, but more so in rural areas and in the north. Broader availability of the pill through pharmacy outlets and of the injection and IUD in stationary health facilities is likely to raise contraceptive use as well. Pharmacies are a particularly good outlet for nonsur- gical methods, as pharmacies are generally more accessible to rural women and unmarried men and women than are health facilities that offer family planning. Furthermore, they may' require less waiting time for the client and may be less conspicuous. The results for the effect of service quality variables are not strong. Of course, the facility questionnaires are not able to capture all of the quality aspects of a service. Interviews with clients and potential clients and observation of service delivery may yield more subjective insights. The Population Council recently conducted an in-depth situation analysis of 147 familyv planning service deliverv points in six states of Nigeria, which included observation of client-provider interactions and interviews with staff and clients (Mensch and others 1994). Multivariate analysis of the number of new family planning clients per 1,000 women resident in the area found no statistically significant effect of the qualitv of provider services. However, the study was not able to control for the socio- economic characteristics of the female clients or for the availability of services, as the present article has done. The final report for the situation analysis noted that queuing is not common at health facilities and method stockouts are not "a major problem" (Adewuy'i and others n.d.). Despite improvements that could be made in sanitation, counseling, and availability of basic equipment, client 184 THL WVORL[) BANK F(ONONMI( REVIEW. VVOL. I, NO) I satisfaction was very high. However, insistence on spousal consent may be lim- iting access to services in many facilities. In a country such as Nigeria, where availability of services and female schooling are both very low, it is difficult to believe that improved service quality (and the implied higher unit costs) of pro- viding it will be the most cost-effective means of raising contraceptive use. Some of the most measurable aspects of quality and availability that public authorities can monitor are measured in the NDHS. Although this study has shown some statistically significant marginal ef- fects of both individual and service factors on contraceptive use, it would be incorrect to infer that these marginal effects reflect the probable outcomes of large increases in the share of women who are educated or in the avail- ability of contraception. The experience of countries such as Zimbabwe shows that substantially raising the schooling of large numbers of women and their access to family planning services (as was done during the 1980s) can have a much greater impact than these marginal effects would imply. Furthermore, we would expect that as contraceptive services become more available in Nigeria, distance will cease to be a binding constraint to increased contra- ceptive use. We have had to assume that the placement of health facilities and pharmacies and the availability of family planning within them is independent of the women's fertility desires or their demand for contraception. Since Nigeria is a country where family planning services are not widely available, this may not be a good assumption. Certainly in the case of private pharmacies, it is likely that the probability of family planning being offered will be influenced by perceptions of the demand environment. We have no way of correcting for the potential endogeneity of the supply of family planning within the existing data set. More accurate estimates of the potential impact of improved availability or quality of contraceptive services could be obtained through a study in which facilities were randomly assigned family planning services and contraceptive use could be ob- served over time. Since one objective of the population program in Nigeria is to improve the availability of methods, it seems to us that careful data collection from women and facilities, as they are sequentially provided these services, would yield even greater insight on the impact of availability and quality of services on contraceptive use. Feviset.zn and Ainsivorth 185 Table A-1. Variable Means for Characteristics of Wom7en, the Nearest Health Facilities, and the Nearest Pharnacies by Various Specifications Specificat on ' Variable 1. 2, 5, 8 3 4 6 7 Dependent variable Current use of contraception 0.029 1.1)29 0.0/52 0.014 0.048 Independent variable Individual wZolman Age (years) 28.6 28.6 28.2 28.6 28.6 Age squared 906 906 87-6 902 91( Urbani residenceb 0.107 0.0()7 (0.402 0.038 0.191 Years of schooling 2.63 2.63 3.72 0.904 4.72 Interaction: urban x schooling 0.788 0.788 2.43 0.104 1.35 Northeast residence" 0.317 0.317 0.2.33 Northwest residence' 0.232 0.232 0.193 Southwest residence5 (0.162 0.162 0.310 Interaction variable Northeast x schooling 0.278 (.278 (.234 Northwest x schooling 0.2 18 (0.218 0.24.5 Southwest x schooling 0.886 0.886 2.(0 Type of floor' 0.529 0.529 0.58 3 0.398 0.688 Protestant' 0.263 0.263 0.335 ).089 0.47.5 Muslimb (.559 0.559 0.487 0.887 0.161 Partner or husband" 0.837 Partner's years of schoolinig 2.27 Nearest facilityv Distance (miles) 5.14 5.73 3.18 7.05 2.83 Outpatient registration fee, 2.33 I.X0 0.993 1.39 3.46 Family planning offered" 0.643 0.747 0.5 18 Injection offered' ( ) .419 0 .526 0.298 0.482 0.342 Injection pricee. 2.72 I . 1.68 2.63 2.82 IUD offered' 0.463 0.5 19 0.394 ILI) price',' 3.04 2.69 3.46 Pill offered' 0.588 (0.735 i).411 (0.663 0.496 Pill price' 1.03 0.787 0.447 0.826 1.29 Condom offered" 0.546 (0.682 0.381 0.701 0.358 Condom pricec,, 0.488 0.611 0.323 0.577 0.380 At least one doctor" 0.559 0.573 0.316 0.521 0.606 Hospital' 0.337 0.413 0.229 (0.4()1 0.258 Health clinic' 0. )79 O1.224 (0.127 (0.266 0.517 Privately owned facilityv 0.212 0.143 0.080 0.075 0.379 Health facility with family planninig linked to the ctiuster ().58 1 0.304 Nearest pharmacy Pharmacy interviewed" 0.733 0.733 (0.437 Distance (miles) 3.03 3.0 1.87 Number of hours open per week 57.4 S7.4 34.3 Family planning offered 0.227 ().227 (0.133 (Tablc crontinues on tbe following page.) 186 THE WORLD BANK ECONOMIC RKVIIW. VOL. 1(, NO. I Table A-1. (continued) Specification' Variable 1, 2, 5, 8 3 4 6 7 Number of family planning methods offered 0.707 (.707 0.294 Injection offered' 0.061 0.061 0.040 Injection pricec.f 0.833 0.833 0.506 Pill offered" 0.159 0.159 0.091 Pill price' 1.12 1.12 0.616 Condons offered' 0.194 0.194 (.108 Condom price'.f 0.360 0.360 0.198 Sample size 4,589 4,589 8,761 2,071 2,518 a. The specificationis correspond to those in tables 5 and 7. b. Dumnmy variable: the value is I if the condition is true; (0 otherwise. c. DuLmImy variable: the value is I if the floor is parquet, vinyl, ceramic, or cement; 0 otherwise. d. For specifications 3 and 4, this is the nearest source of family planning; for all of the other six specifications, it is the niearest health facility. e. Fees and prices are in naira. The average 1990 m1arket exchanige rare was 12.8 naira per L.S. dollar. f. The variable is availability of the method (I or 0) timiies the price. Source: 1990 N1)-IS data. REFER ENCES The word "processed" describes informally reproduced works that may not be com- monly available through library systems. Adewuyi, Alfred, H. Ajaegbu, G. Ayoola, S. Babalola, G. Esimai, M. Kisseka, and K. Omideyi. N.d. Nigeria: The Family Planning Situation Analysis Study. USAIL) Con- tract No. DPE-3030-Z-00-8065-00. Primary Health Care Unit; Federal Ministry of Health; Family Health Services Project; Obafemi Awolowo Ulniversity, Operations Research Unit and Network; and The Population Council, Africa 0R/TA Project. Lagos, Nigeria: Federal Ministry of Health. Ainsworth, Martha, Kathleen Beegle, and Andrew Nyamete. 1995. The Inipact of Fe- male Schooling on Fertility and Contraceptive Use: A Study of Fourteen Sub- Saharan Countries. LSMS Working Paper 110. Washington, D.C.: World Bank. Arowolo, 0. 0. 1984. "Fertility in Nigeria." Paper presented at the Seminar on Law and Population, University of Lagos, Lagos, Nigeria. Processed. Becker, Gary. 1960. "An Economic Analysis of Fertility." In Universities-National Bureau Committee for Economic Research, ed., Demographic and Economic Chanige in Developed Countries. Princeton, N.J.: Princetonl lUniversity Press. 1981. A Treatise on the Family. Cambridge, Mass.: Harvard University Press. Caldwell, John C., 1. 0. Orubuloye, and Pat Caldwell. 1992. "Fertility Decline in Af- rica: A New Type of Transition?" PopulJation and Developmnent Review 18(2):21 1- 42. Federal Office of Statistics. 1968. Rural Demographic Sample Survey 1965/66. Lagos, Nigeria. Fevisetan anzd Ains2vortib 187 Federal Office of Statistics and IRD (Institute for Resource Development)/Macro Inter- national. 1992. Nigeria Demographic and Health Survey 1990. L-agos, Nigeria, and Columbia, Md. Federal Republic of Nigeria. 1975. Third National Development PlaLn 1975-1980. Lagos: Federal Ministry of Economic Development, Central Planninig Office. 1988. National Policy o n Population for Development, Utnity, Progress and Self-Reliance. Lagos: Department of Populationl Activities. Fevisetan, Bamikale J., and Martha Ainsworthi. I 994. Contraceptive Use and the Qual- ity, Price, anid Availability of Fam7lily Planning in Nigeria. LSNIS Working Paper 108. Washington, D.C.: World Bank. Mensch, Barbara, Andrew Fisher, Ian Askew, and Ayorinde Ajavi. 1994. "Using Situa- tion Analysis Data to Assess the Functioning of Family Planniniig Clinics in Nigeria, Tanzania, and Zimbabwe." St udlies in Fanily Planning 25(1):18-31. National Population Bureau (Nigeria), and World Fertility Suirvey. 1984. The Nigeria Fertility Survey 1981/82 Principal Report, 1. Lagos. National Research Council. 1993. Demiographic Elff'ets of Econonmic Reversals in Susb- Saharan Africa. Washington, D.C.: National Academily Press. Pitt, Mark M., Mlark R. Rosenzweig, and Donia M. Gibbons. 1 993. "The Determi- nants and Consequenices of the Placement of Government Programs in Indonesia." The World Bank Economic Revieu 7(3):0 19-48. Rosenzweig, Mark, and T. Paul Schultz. 198 S. "The Supply and Dlemand for Births: Fertility and Its Life-Cycle Consequences." American Economic Review 75(5):992- 1015. Rosenzweig, Mark R., and Kenneth 1. Wolpin. 1988. "Migration Selectivity and the Effects of Public Programs." journal of Puiblic Economics 37:265-89. Scribner, Susan. 1995. Policies Affecting Fertility and Contraceptile Use: An Assess- ment of Twelve Sub-Saharan Counltries. World Bank Discussioll Paper 259. Wash- ington, D.C. Thomas, Duncan, and Johin Maluccio. 1995. Contraceptive Choice, Fertility, and Pub- lic Policy in Zimhbabi'. I_SMS Workinig Paper 109. Washington. D.C.: World Bank. UNICEF (United Nations Children's Fulid). 1985. The State of the World's Children. New York: Oxford tJniversity Press. United Nations. 1985. World Population Prospects: Estimazltes as Assessed in 1982. New York. World Bank. 1992. Worldl Development Report 199): Development and the Environ- ment. New York: Oxford University Press. Fertility, Contraceptive Choice, and Public Policy in Zimbabwe Duncan Thomas and John Maluccio Zimbabiwe has inviested massively in ptiblic infrastructure since inidependence in 1 980. The impact of these inv!estments on demiographic outcom}es is examined uisintg household survey data miatched with tufo community level surveys. A w'omnan 's edii- cation is a powferful predictor of both fertility and contraceptive use. These relationt- ships are far from linear and havie changed shape in recent years. After conztrolling for household resources, both the availability and quality of bealth and family plani- ning services have an iniportant impact on thle adoption of miodern contracepti ves. In particular, outreach programs such as mobile fnamily planning clinics and comn- munity-based distributors (CBDs) habve beeni especially successhfl. Houwev'er, niot all womenc are equally served by this infrastructire. For example, (;IiS have a bigger impact on youinger, letter educated ivonien. iwbile moobile famnly plaznzinzg cliniics appear to have more success with older, less educated women. Since independence in 1980 the government of Zimbabwe has invested mas- sively in infrastructure and a large share of the public budget has been allocated to the provisioni of social services, particularlvl health and education. For ex- ample, between 1980 and 1986, enrollment ratios rose by 40 percent among primary school age children and almost sixfold (from 8 to 46 percent) among those of secondary school age. Today the vast majority of the population has access to primary education. The family planning program, whichi has been inte- grated into the public health system since the mid-1980s, has expanded dra- matically since independence (Boohene and Dow 1987), and there has been con- siderable effort to provide services to the poorest Zimbabweanis. Duncan Thomas is with the Labor and Populationi Programii at RAND aind the EconoIm1ics Departmenlt at 1ICLA, and John Maluccio is in the Economics Department at Yale University. This article was written as backgrounid for the researclh project on1 '"The Economilc and Policy Determiinants of Fertility in Sub- Saharan Africa," financed by the World Bank Research Coommittee (RP0 6o1 69 1 6 and sponsored by the Africa Technical Departmiient and the Policy Reseac-h D)partment of the World Bank. The atithors gratefully acknowledge finanicial support froTm the Diis Smiall Grants Prograimi ftinded by the Andrew Mellon Foundation, a National Science Foundatiorm Gradkuate Research Fellowship, the World B1ank, and Yale Ulniversity Center for International Area Stidies. The author-s thalnk the Institute for Resource Development (iRD)/Macro Internatioiial, the Central Statistical Office in Harare, Zimbabwe, the World Bank, and the Zimbabwe National Family Planning Council for makinig data available. TIhe authiors appreciate the conimients of Martha Ainswortih, Susan ( ochraile, EliLaheth Frankenberg, Charlie Griiffin, Pelad Namfua, and two anolnymous referees, and the guidlance of ltyai NILuvandi. 3 1996 The Internationial Bank for Reconstructiol and [Developiment / 11 WORLL) BANK 1 89 1 90 THE WORID BANK LCONOMI(: KEVIFW, VOL. 10, NO. I What have these investments bought in terms of demographic outcomes? Among Sub-Saharan African countries, Zimbabwe, along with Botswana and South Africa, leads the pack in terms of adoption of modern contraceptive meth- ods. According to the 1988 Zimbabwe Demographic and Health Survey (ZDHS), 27 percent of women age 15 to 49 were using a modern method at that time, almost 50 percent had used a method, and knowledge of modern methods was virtually universal (cso and IRD/Macro International 1989). Prevalence rates are roughly similar in Botswana (where per capita gross national product (GNP) is about 50 percent higher than in Zimbabwe) and these rates are about twice those reported in Kenya. Prevalence rates in the rest of continental Africa (apart from South Africa) are typically below 10 percent (National Research Council 1993a). Yet, in spite of the apparent success of family planning in Zimbabwe, fertility remains high: the total fertility rate (TFR) was estimated to be about 5.5 in 1988 (cso and IRD/Macro International 1989). Critics have charged that the family planning program is not effective. Mauldin and Ross (1991), for example, rank the strength of the program as "moderate" and well behind that of Botswana. Some studies have suggested that a majority of women use contraceptives for birth spacing rather than stopping (Way, Cross, and Kumar 1987); others have suggested that contraceptives are used inefficiently (Adamchak and Mbizvo 1990); and the family planning program has been crit- icized for relying too heavily on only one modern method, the pill. There has not, however, been any systematic evaluation of the impact of the public investments in health and education infrastructure in Zimbabwe over the last decade. Because the effects of these investments are likely to be greatest on the most recent age cohorts, it will be decades before it is possible to provide a complete and definitive answer to the question of what the investments have bought. But policy decisions cannot wait decades and some aspects of the ques- tion can be addressed. This article attempts to do just that. Using microlevel data, we begin with an examination of the determinants of fertility outcomes in Zimbabwe, focusing on the role of the household and its resources, in particular education and measures of income. To assess the impact on fertility of public investments in health and family planning services, histori- cal data on those services are needed because fertility reflects the cumulation of choices, attitudes, and service availability over a woman's entire childbearing life. This is true even of current fertility (such as births in the last five years). Although the surveys we use are rich, they do not contain historical information that would allow us to directly examine the link between infrastructure and fertility. We turn, therefore, to a choice related to fertility outcomes, the deci- sion to use modern contraceptives. In addition to examining the influence of household resources, we place the spotlight on the role of the availability and quality of community health and family planning services. We pay special atten- tion to the distributional impact of investments in health programs bv exam- ining differences in the effects of the programs across educational groups and different cohorts of women. Thonmas and Maluccio 191 Consistent with much of the rest of the literature, we find that education is a powerful predictor of both fertility and contraceptive usage. Moreover, these relationships are far from linear and have changed shape in recent years in Zim- babwe. For example, among women age thirty-five and above, there is no signifi- cant relationship between education and the number of children ever born; among younger women, however, the relationship is negative and significant, and even women with relatively little education have fewer children than their less edu- cated peers. Turning to community characteristics, the results indicate that health and family planning services do have a positive impact on adoption of modern methods, after controlling for household resources. In particular, outreach pro- grams such as mobile family planning clinics and community-based distributors (C(BDs) have been especially successful. Not all women are equally served by this infrastructure. For example, CBDs have a bigger impact on younger, better edu- cated women and mobile family planning clinics appear to have more success with older, less educated women. The next section outlines the conceptual model guiding the empirical analy- sis. Section 11 describes the data. The results are presented in section 111, and a conclusion follows. 1. MODEL Following the economic model of household production (Becker 1981) ap- plied to fertility and contraception (Rosenzweig and Schultz 1985), we assume that households choose to allocate resources in order to maximize utility, which depends on consumption of market and nonmarket goods. Consumption in- cludes the leisure of all household members as well as, among others, the quan- tity and quality of children. Household choices are made based on a budget constraint and the technology underlying home production, including the produc- tion of child quantity and quality. On the one hand children are valued in and of themselves, and may also be viewed as productive assets, providing labor while young and yielding returns in terms of income support to parents later in life. On the other hand children impose a time and money cost on the household. A desire to lower fertility below its natural (or unregulated) level will depend on the pecuniary costs of raising children as well as on the imputed value of time costs and thus the value of parents' time. Hence, desired fertility will be lower among couples who earn higher wages and who are better educated (conditional on income). If women bear the brunt of childrearing, then the impact of mater- nal education is likely to be larger than that of husbands. But even if hus- bands spend no time raising children, as long as male and female leisure times are complementary, the husband's education, like the wife's, will be negatively associated with fertility. Income effects are ambiguous. If children are normal goods, higher income will be associated with more children. But if demand for child quality rises with 1 9 2 THE WoRI D BANK iCONONl: RVIMEW, VoL. D!. No. I income, this relation will operate in the opposite direction and confound the income effect. Moreover, if our income measures do not adequately measure long-run household resources, then they may be proxied by husband's educa- tion, which muddies the interpretation of that covariate. Focusing on these indicators of household resources, we first estimate the reduced form demand for children, N: (1) N = N(Pb' which depends on household characteristics, ph' and unobserved household- level heterogeneity co,,, representing, for example, differences in fecundity and tastes for children, which are assumed to be random and uncorrelated with household characteristics. If the costs of children outweigh their benefits, couples may choose to limit the number of pregnancies and thereby control the quantity of children. One method of limitation is the adoption of contraceptives. Since there is evidence that short birth intervals are associated with poorer maternal and child health outcomes, contraceptioni may also be used for birth spacing. Of course, wider birth intervals for any woman will also be associated with completed fertility (of women forty-five and older) being below its natural level. The decision to use contraceptives will depend on their perceived costs and benefits. Costs will depend on the efficacy of the contraceptive, its price, and the difficulty of obtaining and using it. Thus contraceptive usage is likely to be af- fected by supply-side factors. For many women, a large element of the full price will be captured by the availability of contraceptives. A key aim of this article is to examine the influence of a broad array of community level indicators of the availability of services on contraceptive use. We pay special attention to service quality and are particularly interested in the impact of the community-based distribution mechanism that has been in operation in Zimbabwe since indepen- dence in 1980. The benefits of contraceptive usage will be associated with the fecundability of the couple and also their desire to reduce fertility below its natural level. The benefits will vary with age, and this relationship is likely to be nonmonotonic. Very young women may adopt contraceptive methods to delay childbearing. As they move into a childbearing period, usage will decline only to rise again when they decide it is time to stop bearing children. Things are a bit more complicated if contraceptives are Llsed for birth spacing, because women in their twenties and thirties are also likely to be contraceptive adopters. Finally, since fecundability tends to be lower among older women, the benefits of contraception will not be as great and so usage is likely to be lower among these women. If, as argued earlier, fertility is lower among higher wage and thus better educated womiien, contraceptive usage will tend to rise with maternal education. Of course, educationi is likely to play many roles, over and above its influence on the value of time, including improved information processing and possibly effi- ciency in usage of contraceptives. In addition to its impact through fertility, Thonzas and Maluccio 193 income may affect adoption of contraceptives if they involve resources, in which case the poorest may be less willing (or able) to buy contraceptives than the better off. Our second empirical model is a contraceptive usage, ir, function: (2) 7r = Xf (Ph, Po E) where 1tc are community level characteristics. Unobserved household-specific heterogeneity, such as fecundity and efficacy of use, is captured in Eh, which is assumed to be random and uncorrelated with either p,, or u,u. We thus rule out migration of women or couples to places where services are better in order to obtain access to contraceptives; endogenous program place- ment is also ruled out (Rosenzweig and Wolpin 1982, 1986). In the case of Zimbabwe, these assumptions do not seem to be obviously foolish. There are no taboos associated with using contraceptives, which are relatively readily avail- able. Migration, it seems, is likely to be motivated by more than a desire for contraceptives. Since 1980 many clinics and several hospitals have been opened as public health policy has been seeking to bring basic services to the ma jority of the population. Because increasing contraceptive usage has not been the highest priority in public health policy, the placement of these programs is likely to have more to do with health problems and inadequate health services in the vicinity than contraception per se. Some evidence is presented that suggests that endogeneity of program placement may not be critical in this context, at least for large-scale investments. Nevertheless, it is not possible to definitively test these assumptions with the data at hand, and thus the conclusions regarding the impact of community characteristics need to be interpreted with this caveat in mind. Some authors have suggested using first-difference estimates to control for the endogeneity of program placement (Pitt, Rosenzweig, and Gibbons 1993; Gertler and Molyneaux 1993; Frankenberg 1995). We explore a related strat- egy later in the article. Another potentially serious issue is that community services tend not to be located in the poorest regions of the country. Failure to control for household, or at least community, income levels could thus lead to substantially incorrect inferences. For example, suppose that contraceptives are more likely to be used by higher-income women who tend to live in areas that have better health ser- vices. Without good controls for resources, an empirical correlation between service quality and contraceptive usage does not have an unambiguous interpre- tation: it may simply reflect the impact of the omitted covariate (income) or it may in fact be that service quality does affect usage. The survey data, discussed more fully later, contain limited information on household resources. All regressions, hut one, include controls for the household's possession of a range of assets. Income of men is nor reported, and so the regres- sions include a good predictor of it, the husband's education. As discussed above, his education will thus capture both a value-of-time effect and an income effect. Because not all women are married, controls for marital status are also included 194 THF. WORI D BANK ECONOMI( RFVIFW. VOi . 10, NO. I in the model. It could be argued that marital status should be treated as endog- enous in a model of contraceptive choice. But, given that a key contribution of this article lies in the examination of how contraceptive choice is influenced by community services, we choose to ignore that potential problem in an effort to control for income to the fullest extent possible. Estimates that exclude marital status and husband's education are discussed in the text; these and other esti- mates not presented here are reported in Thomas and Maluccio (1995). 11. DATA AND METHODS Data are drawn from three sources: a household survey and two specially conducted community level surveys, which we describe in that order. The ZDHS, conducted in 1988 by the Central Statistical Office (cso) in collabo- ration with the Institute for Resource Development (IRD)/Macro International, Inc., is a nationwide survey that interviewed over 4,200 women and their fami- lies in 166 clusters. The survey provides detailed information about the fertility history of these women, their contraceptive use and knowledge, their health, and that of their children. The survey respondents also provided limited infor- mation about the socioeconomic status of their households, including their own education, that of their husband, and household's ownership of any of a series of assets. The data are fully described in cso and IRD/Macro International (1989); key summary statistics are presented in appendix tables A-I and A-2. The ZDHS is part of a worldwide program that has been operating since the mid-1980s and has successfully fielded surveys in a wide array of countries. These surveys have provided an extremely rich data source, which has led to substantial contributions to the understanding of demographic change in many countries around the world. For an excellent discussion focusing on Africa, see the series of volumes published as part of a National Academy of Sciences study (National Research Council 1993a, 1993b, and 1993c). According to the ZDHS, the average Zimbabwean woman has given birth to nearly three children, and the number increases with age. Among women age thirty-five and older, the average is 6.2, and completed fertility (of women forty- five and older) is 6.8 children; it remains to be seen whether completed fertility among the current cohort of women in their twenties and thirties will be as high as the cohort of women in their late forties. Contraceptive knowledge is virtu- ally universal, and 27 percent of the sample women report they are currently using a modern method. For over 87 percent of these women, that method is the pill; 6 percent are sterilized, 3 percent use condoms, and 3 percent (or twenty- nine women) use intrauterine devices (lUDs). Traditional methods are very un- common, especially among younger and urbani women: less than 5 percent of the sample womeni report they are currently using a traditional method. The overwhelming domiiance of the pill reflects policies during both the pre- independence government (which tended to emphasize birth spacing) and the first post-independence government (which banned the injectable Depo Provera Thomas and Maluccio 195 in the early 1980s). In more recent years there has been a concerted attempt to expand the choices available to women. Because very few women in the survey use a modern method other than the pill, most of the analyses reported later do not try to distinguish among the methods. About 14 percent of women in the survey report having no education and about half have completed primary school or more, although nearly half of those women exited at the end of primary school (seven years of schooling).' These averages mask substantial differences across cohorts, because the youngest have benefited from heavy investments in education since independence. For example, among women thirty-five and above, only 11 percent had gone beyond primary school, but among those who are less than thirty-five years old this proportion is almost 40 percent. Although about two-thirds of the sample women are currently married, less than three-quarters of married women are living with their husbands; split households are common in Zimbabwe, and in the majority (but not all) of the cases the woman lives in the rural area while her husband works in a town. These household level data have been matched with two community level data sources. First, in 1989-90 the cso and IRD/Macro International resurveyed the same 166 clusters covered in the ZDHS and obtained extremely detailed in- formation about general infrastructure, along with the availability and quality of health and family planning services in the community. Those data, recorded in the Zimbabwe Services Availability Survey (cso and IRD/Macro International 1991; see also Wilkinson 1992) can be directly matched, at the cluster level, with the ZDHS. Information about local services was gathered from "knowledgeable" com- munitv informants who were typically identified by the village head; the num- ber of informants in each community ranged from two or three to twenty or more. In addition to describing services in the vicinity, the informants were asked about six types of facilities-general or district hospitals, rural hospitals, clinics or health centers, Zimbabwe National Family Planning Council (ZNFPC) clinics, pharmacies, and private doctors. They were asked to identify the nearest facility of each type and the distance to it. If the facility was within 30 kilometers of the community center, it was visited by an enumerator who collected information on the nature, quality, and prices of services offered. The collectioii of community level data is far from trivial, and it is not obvi- ous how best to identify local providers or, put another way, the appropriate catchment area for a particular provider. Unfortunately, it appears that infor- mants in several communities were not very well informed. For example, infor- 1. Years of schoolinig is constructed from responses to questions about completed level and the grade within that level. For primary school, women were able to list op to five grades, so completion of Standard 5 (seven years of schooling) is treated a1s completion of primary school. For secondary school, completion of Form 2 is treated as equivalent to nine years of completed schooling, and Form 4, eleven years. The small fraction of woomen who have twelve or more years of schooling are grouped together (see appendix table A- I). Also, we nlote that twelve percent of the womenl in the survey have not yet completed schooliig. 1 96 rHF WXORL[) LANK l (ONONMI( RlVIA:M. Vol. I,,, NO. I mants in 8 percent of the clusters failed to identify a general or district hospital.2 Exactly why a hospital was not identified is unclear, and one may argue that if the informant could not identify it, this may reflect the general perception of people in the community that they do not have access to a hospital. Along the same lines, but perhaps a bit more problematic, is that ZNFPC clin- ics were identified by informanits in only 16 percent of the clusters. In part this reflects the fact that these clinics do not cover the entire country and are concen- trated in towns and cities, so that rural informants may not even be aware of the clinics. But in Harare, for example, where there are seventeen clusters, the ZNFPIC clinic was identified by informants in only one of those clusters. Since the ZNFPC clinic shares the same grounds as the Harare Hospital (but is a completely independent entity and located aparr from the hospital), we can deduce the distance to the ZNFPC clinic using reported distances to Harare Hospital: in every case, it is less than 30 kilometers. This calls for caution in relying on information about only facilities identified as being within a 30 km radius of the community (and thus visited by the enumerators). For the purposes of family planning ser- vices, therefore, we will define the community as the district (there are fifty-two districts). Because over the last decade public policy in Zimbabwe has shifted dramati- cally toward the provision of family plannling services through CBDs, it is of considerable interest to look closely at C:BDS' role in affecting women's choice to use modern methods. Information on CBDs being rather limited in the Services Availability Survey, we turn to a second special community survey, the Zimba- bwe Situation Analysis Study. The Situation Analysis Study was conducted in 1992 by the ZNFPC, the Popu- lation Council's Africa Operations Research and Technical Assistance project, and the Family Planning Service Expansion and Technical Support project (ZNFPC 1992). The objective of the Situation Analysis Study was to provide comprehen- sive information about the availability, functioning, and quality of family plan- ning services in Zimbabwe. In addition to detailed information on individuals who are CBDs, the survey provides data on family planning services available from (public and private) clinics. Information was collected in structured inter- views with providers and clients (includinig exit polls), as well as direct observa- tion of clinic conditionis and provider-client interactions. The sample design, which is described in detail in ZNFP; (1992), is essentially facility based. We have, therefore, matclhed the facilities with the household data at the district level. It should be noted that not all districts in Zimbabwe were included in the ZDHS and thlat for those which were included but do not have data in the Situa- tion Analysis Study, we have matched data from a neighboring district. The Situation Analysis Study gatlhered information from 181 clinics and 140 CBDS. 2. One of these clUsters is in the Jambesi area of the Hwange District, and the informants failed to identify Hwanige Hospital; in the nieighboring district to the sooth, Gwaai, Hwvange was identified by the isforiants, bUt it was too far away (8-7 kilometers) to be selected tor a visit. Thomas and Maluccio 1 97 Information about CBDs and clinics contained in the Services Availability Survey and Situation Analysis Study overlaps. It is thus possible to cross-check the two sources. Because the surveys were conducted two to three years apart, the data contained in them are unlikely to be identical. Good correspondence between them, however, would suggest that district level matching of the two community surveys is reasonable in the Zimbabwe context. Similarity may arise either because there is only one provider of a particular type in each district or because the intradistrict heterogeneity is considerably smaller than the interdistrict variation. It turns out that the information contained in both surveys is remarkably close. In only 8 percent of districts in which the Situationi Analysis Study recorded an interview with a CBD, the Services Availability Survey reported no CBDS in that district. In part, this discrepancy may reflect expansion of the CBD program during the three years between the fielding of the Services Availability Survey and Situation Analysis Study.3 There is also broad agreement about services available in clinics. For example, in only 6 percent of the districts was there disagreement about the presence of state-certified nurses, and in less than 5 percent of districts there were discrepancies in the availability of condoms and oral contraceptives. In general, we find that the intradistrict (and intersurvey) heterogeneity is tiny relative to the interdistrict variation in the data. The results are quite reassuring. They suggest there have not been large changes in the provision of these services during the period between the two surveys. This is not too surprising in view of the fact that real public health expenditures grew little during the late 1980s and early 1990s. This is important because we are implicitly assuming that no shifts occurred in the relative distribution of avail- ability and quality of services provided in Zimbabwe between 1988 (when the ZDHS was fielded) and 1992 (when the Situation Analysis Study was completed). Appendix table A-2 summarizes some of the community level information. The first column is measured at the cluster level; the second to fourth columnis reflect the proportion of women living in communities with the services. Of interest is the fact that about half the sample women live within the vicinity of a general hospital (and virtually all urbani woman are close to a facility of this sort). There is a clinic in every community, and almost half of them have been built since 1980. About two-thirds of the communities are visited by a health worker but only half of those are also served by a mobile family planning clinic. CBDs are operating in about two-thirds of the communities. 111. REGRESSION REStUTS Using the microdata in the ZDHS, the analysis begins with an examination of the impact of household characteristics, and especially a womani's education, on 3. In 20 percent of the districts there was a discrepancy in the gender of the (B); the Situation Analysis Stmdy reported interviews with considerahlv more male CSrOs. This is unlikely to reflect changes in the (BD program and suggests, perhaps, that local informants in the Services Availahility Suirvey may have been misinformed. 198 THE WORLD BANK ECONONMIK REVIEW. V(OL. L1, NO. I fertility. We proceed to assess how the same characteristics affect the probabil- ity that she uses modern contraceptives. To discern the influence of a series of community service characteristics on this probability, we incorporate data collected in the Zimbabwe Services Availability Survey and the Zimbabwe Situ- ation Analysis Study and estimate the reduced form, equation 2. Fertility Fertility in Zimbabwe is high. Table 1 presents evidence on the household factors that affect one measure of fertility, the number of children ever born. Age at first birth and birth spacing are also discussed briefly. The model is esti- mated by the method of least squares, which ignores the fact that the outcome is discrete. The results do not rely on this assumption: a Poisson model of children ever born provides essentially identical results. Each woman's own education is included in the regression with no parametric restrictions placed on the shape of the relation between it and fertility: each year of completed schooling is repre- seinted by a dummy variable (except for education of twelve years and above, which is equivalent to continuing beyond "O" levels and accounts for less than 2 percent of women in the sample; see footnote 1). The coefficients represent the difference in the number of children born to a woman with a given educa- tion level and to a woman with no education, holding other background charac- teristics constant. In the first column of table 1, education of the woman is included along with her age (represented by dummies for each five-year age group), marital status, ethnicity, and sector of residence.4 Some authors, including Cochrane (1983) and Cochrane and Farid (1990), have noted that in many countries the impact of maternal educa- tion on fertility is not monotonic. For women with very little education, the number of children ever born rends to rise with education until some threshold level (typi- cally somewhere between five and seven years of schooling) is reached; thereafter, fertility and education are negatively correlated. A similar pattern is observed in Zimbabwe. The first few vears of primary schooling are unrelated to fertility and it is only when women are close to completing primary school (specifically, six years of education) that there is a significant negative association between education and fertility. This correlation is not constant but tends to increase with educa- tion, particularly at the top of the distribution. To what extent can the effect of education be explained by income? We at- tempt to address this question in the second column of table 1, in which husband's education and a set of dummies for ownership of a range of household assets are added to the regression to control for permanent income and wealth. Income effects are small, and hLisband's education only matters among the better edu- 4. Dropping controls for marital status restlts in slightix lower education effects but only one of these differences is significaant (eleven years of schooling). The X- sratisric for significance of all differences in educationi effects is 5.4 with a p-value of 0.94. Endogetneitv of marital status seerns to he of second-order importance in its effect on estimated educationl coefficienits in these regressionis. The same conclusion applies to the contraceptiVe Luse regressioiins diScussed later. Thomas and Maluccio 1 99 Table 1. Number of Children Ever Born: Role of Educationz and lincomne Woman's education Womant's and By wvomnan's age education household Less than 35 or B-v womnan's residence Variable only inconie S older Urban Rural Woman's education (completed years)'- 1 -0.193 -0.191 -0.545 0.410 0.676 -0.369 (0.88) (0.88) (2.8 (0.78) (1.41) (1.48) 2 (complete preschool) 0.005 0.028 -0.1 17 (i.25(0 0.104 0.024 (0.04) (0.18) (0.85) (0.73) (0.33) (0.14) 3 0.089 0.105 (.087 0.287 0.391 0.046 (0.63) (0.74) (0.66) (0.89) (1.37) (0.28) 4 -0.191 -0.1 5 -) .395 0.253 0.067 -0.1 35 (1.37) (1.11) (3.24) ((0.73) (0.27) (0.8() 5 -0.157 -0.096 -0.3 3 0.349 -0.129 -0.062 (1.39) (0.83) (3.30) (1.22) (0.56) (0.47) 6 -0.405 -0.307 -0.4(65 0).144 -0.314 -0.209 (3.50) (2.54) (4.72) (((.42) (1.40) (1.46) 7 (complete primary) -0.5 16 -0.365 -0.461 -(.()061 -0.306 -0.266 (5.32) (3.49) (S¶38) ((0.2(0) (1.66) (2.08) 8 -(0.640 -0.417 - ((. 0.152 -0.4 59 -0.291 (4.26) (2.67) (4.3(1) (0.20) (1.86) (1.47) 9 (complete Form 2) -(0.731 -(04.56 -(1._5' -0 346 -0.442 -0.383 (5.91) (3.40) (5.49) ()07) 2 .15) (2.12) 10 -(.793 -0.528 -0.744 -0.308 -).611 -0.477 (4.. 1) (2.90) (5.60)) ()0.29) (2.40)) (1.88) 11 (complete Form 4) -1.051 -0.697 -0.959 -(0.49 1 -(.7 .777 (8.81) (5.18) (9.33) ((0.7) )3.64) (4.)4) 12 or more -1.418 -).974 -1.364 -0.611 -1.211 0).056 (6.39) (4.17) (7.78) ().77) (4.55) (0).1I0) Husband's education (completed years), Complete preschool 0.112 (0.268 -0.042 -0.814 0.203 (0.54) (1.49) (0.08) (1.63) (((.86) Some primary school 0.168 0.384 -0.111 0.228 0).136 (1.70) (4.28) (0.49) (1.19) (1.18) Complete primary school -0.013 0.11() -()(.)51 -0.245 () 11() (0.15) ( .32) ().20) (1.40) (0.92) Complete Form 2 -0.199 -0.141 ).131 -0.147 -0).194 (1.64) (1.45i (0.37) (((.78) 1(.23) More than Form 2 -((.522 -(0.246 -0.3 16 -0.467 -0.466 (4.33) ( .(.2) (().74) (2.58) (2.78) (Table continues on the tollowing page.) 200 THF WORI D RANK V.LONOMI( RlV[lW. VOl 1(), NO. I Table 1. (continzued) Womtan 's education Wo,nanz s an1d By w'onan's aige ediucation bhousehold Less than 3S or hB womlan s residence Variable only invcomne .3 oldel- Urban Rural Houisebold assets' Motorcycle or bicycle 0.061 -0.(11 0.24() 0.096 0.047 0(.98) (0.24) (1.27i (1.07) ((0.59) (Car -0.288 -0.062 -1.320 -0.304 -0.240 (2.92) ((.8X5) (3.92) (2.95) (1.35) Radio -0.066 -0.().3 -0).06 -0.068 -0.025 (1.06) ((.70) (0.18) (0.69) (0.33) Television 0.076 0. 02) 8 0.0.58 (0.20(1 -0.263 ((.75) (0.37) (0. 16 (2. 03) ( 1.04) Cattle -0.053 -0.039 -(.174 0.188 -0.086 (0. 74) (0.68) ()0.80 (0.68) (1.06) Goats or sheep (.077 0.064 0.081 -0.253 0.087 (1.06: i. I I) ((.38) (0.8X5) (I.0)X Rural residence' 0.393 O).286 ()I.2 ().909( (6.24) (3.53) (2.))) (3.40)) Goodeness o/fit X2 (own educationi) 10.65 3.78 0.97 0.44 3.91 1.92 0(.00 1()0(0l 1().(( [(.9-5] [1(()l 10.031 x2(husbaid's education) 38.98 7.57 11.1' 0.26 .3.90 3.63 10.0(] [0.()() I 1(.001 10.931 10.(] [0.00( F (ill covariates) .362.14 249.2() 218.17 8.61 78.00 1'77.81 1(0.(1 1(.(01 1(.(0(1 1 .1 ) I).(( 1(0.(01 R2 (0.67 (0.67 (.(69 ().20 0.65 0.68 NLumnber of observationis 4,201 4,72)1 3.,129 1,072 1,407 2,794 Note: The values are ordinary least squares (ots) estimates. The regressiois also iniclude conitrols for age (represented by dumminies for each five-year age grouip), mlarital status, and presenIce of huisbland in houselhold. t-statistics are in parentheses, and p-values arn iln square brackets. a. Dummiiv variable: value is I if condition is trLe; 0 otherswise. b. The coefficients represent the difference in the numibher- of childr-en horn to a woman wyith a given education level and a vomalnn with no eduIcation, holding otlerl backgrOUnd characteristics constant. Source: Authors' c.alcUlations based on cs:so and iRiVMacro Internatinonal ( 1989). cated. Nevertheless, a good part of the effect of female education does appear to operate through income and the estimated effects decline from between one- quarter and one-half; this decline is greatest among the better educated. In order to take account of all unobserved community level heterogeneity, including in- come differences, we have also estimated the miodels with cominiullty fixed ef- fects. The educationi effects are very similar to those reported in the second column. For example, the fixed-effects estimates indicate that women withi twelve or more years of schooling have 0.972 fewer children than those with no school- ing; the estimate without fixed effects is 0.974. Thomas and Maluccio 201 Although these regressions control for the age of the woman, it is possible that the effects of education vary with age because the number of children ever born rises with age and there have been dramatic increases in educational at- tainment among recent cohorts in Zimbabwe.5 The third and fourth columns of table 1 present separate regressions for younger women (under thirty-five years old) and older women (thirty-five years old and above). Among older women, there is no significant relationship between education and fertility. Among younger women, however, even fairly low levels of pri- mary schooling are associated with lower fertility. This suggests that invest- ments in education over the last decade are likely to have a substantial payoff in terms of reduced fertility in coming decades. The powerful negative association between education and fertility is largely an urban phenomenon (fifth column). Since younger, urban women are more likely to be participating in the formal labor market than their older or rural counterparts, this suggests that growth in employment opportunities may have a substantial impact on fertility rates in Zimbabwe. This is a hypothesis that warrants careful scrutiny: if true, then it may be that recent declines in labor demand will be associated with a flattening of the fertility-education profile. A very similar pattern emerges for the impact of education on age at first birth (not shown). Women who complete primary school have their first child about seven months later than those who do not (and this difference is significant); women who complete eleven years of schooling wait another two years before having a child. The biggest effect, however, is among women with at least twelve years of schooling: they delay childbirth for five years relative to those Without any schooling. Among older women, there is little evidence of a significant rela- tionship between education and age at first birth except for the tiny fraction with more than ten years of schooling; in fact, relative to women with no schooling, those with less than six years of education tend to give birth at earlier ages. For more recent cohorts, in contrast, the impact of education is positive at all levels of education and significant for women with at least four years of schooling. Apparently part of the negative correlation between education and fertility among recent cohorts can be attributed to delay of the first birth. Like age at first birth, birth space tends to rise with education and the effect is significant only among better educated, younger, and urban women. Counterbal- ancing this trend is a decline in birth space across cohorts: women age forty-five to forty-nine space their children about two months further apart than twenty-five to twenty-nine year olds. Although significant, this difference in birth space is small; so assuming that the more recent cohorts do not continue childbearing longer than their mothers, it seems reasonable to expect the relationship between education and completed fertility to become more powerful in vears to come. 5. In addition to the fact that the impact of public investmenits will differ across cohorts, it is possible that the end of the civil war in 1980 brought with it greater stability in Zimbabwean households which mav have resulted in a baby boom in the early 1 980s. 202 THE WORI.D BANK ECONOMIC REVILW, Vol. 10, No. I Returning to the second column in table 1, the husband's education follows the same inverse U shape reported for women.6 Male education has a signifi- cantly depressing impact on fertility only at the top of the education distribution (nine or more years of schooling) and only among younger women. In fact, conditional on wife's education and income, men who have some primary school- ing tend to have more children than those with no education at all; this effect is large and significant among young women. As discussed earlier, male education effects do not have an unambiguous in- terpretation and are likely to be capturing both the role of income and the value of time. About 30 percent of married women are not living with their husbands; in those households, husband's education should have no value-of-time effect: we find that husband's education is positively, but not significantly, associated with fertility. Among couples living together, however, the husband's education effect is negative and significant if he has completed primary school. (The co- efficient on completed primary school is -0.29 with a t-statistic of 2.0 and on completed Form 2 it is-0.61 with a t-statistic of 3.6.) Recognizing that husband's presence is not exogenous, these results suggest that the effect of his education on demographic outcomes does combine the role of income and the value of time, with the latter being especially important among the better educated. The dummies for ownership of assets also capture income effects. They are small and, apart from cars, insignificant. The effect of car ownership is substan- tially smaller among younger women, and it is only among older women that these income measures are jointly significant. In sum, there is a negative association betweeni education and fertility, albeit among the better educated. The shape of the relationslip appears to have changed among recent cohorts, with education becoming a powerful force behind fertil- ity decline in Zimbabwe. Ulnderstanding the mechanisms underlying the change in this relationship is key for policy design and will be taken up again later. Even after controlling for education and income, rural women tend to have signifi- cantly more children, although this gap may be declining over time. Among older women (thirty-five and above), rural women report about one more birth than those in urban areas, hut among womein less than thirtv-five vears old the difference is only about one-seventh of a child. This is a large decline in the urban-rural gap, although it is obviously impossible to determine at this time whether it will be completely offset by rural women having more children later in life. A natural hypothesis would be that the urban-rural gap reflects differences 6. The models have also been estimated using dumimniies for eachl sear of husband's education (paralleling the female specification). To save space, we report results for the more parsimonious specification because the restrictions it imposes are not rejected and effects of other covariates are essentially unaffected by the choice. Although the main results are captured in the miore parsimiloniious specification, the semiparametric estimates are useful tol draw direct comparisotns between male and female edlucation effects. At the top of the distribution, rhe negative effect of husbanids' educationl is about two-thirds the magnituide of women's education: the coefficients on eleveni years and twelve or m)ore years of schooling are -0.49 and -0.72, respectively, with t-statistics of 3.7 in both .ases. In addition, the inverse U shape of husband's education is more strikiing (and significant). Thomas and Maluccio 203 in infrastructure in the two sectors. Without data on the services available to each woman over her childbearing life, it is impossible to test this empirically. However, contraceptives are used to control fertility, and it is possible to assess whether current infrastructure has any impact on current contraceptive use. We turn now to that question. Contraceptive Usage In the following two subsections, tables 2 and 3 present the estimated effects of household and community characteristics on the probability that a woman is currently using a modern method. Since the dependent variable is dichotomous, we assume errors are distributed as a Gaussian, and estimate a probit model by maximum likelihood. Estimated coefficients have been translated into slopes and are multiplied by 100 in the tables.7 In view of the tiny fraction of women who are using traditional methods, we do not distinguish them in the tables. Estimates based on a trivariate multino- mial logit indicate that few of the characteristics in these data explain the choice to use traditional methods.' Recall that of 27 percent of women who are using modern methods, 24 percent use the pill. We have also estimated quadrivariate multinomial logits (distinguishing women who use the pill, other modern meth- ods, traditional methods, and no method). In our view, the data do not have enough information for us to reliably estimate the determinants of other mod- ern methods, so we discuss only robust results here. Paralleling table 1, we begin with household characteristics in table 2 and then examine the impact of com- munity characteristics on contraceptive use in table 3. Household Characteristics Many studies have demonstrated that better educated women are more likely to adopt contraceptives. Although this is also true in Zimbabwe, the effect is significant only after women have completed several years of pri- mary school, and the relationship is not monotonic (table 2). Among women who did not complete primary school, the function is fairly flat, and these women are about 4 percent more likely to adopt contraceptives than those without any schooling. There is, however, a dramatic step in the function at completion of primary school (seven years of education), when the impact doubles to 8 percent. It is hard to imagine that women learn about the value of contraceptives only in their final year of primary school; instead, this sug- gests that there is something intrinsically different between women who do and do not complete primary school. This interpretation is bolstered by the fact that the estimated effect for women who complete eight years of school- 7. The coefficient estimates, 3, are multiplied by the mean probability, which is the value of the cumulative density function of the normal distribution evaluated at the sample mean, )(xX3 ). The tables report 1001 £b(x1) 8. Older women and women in households with cattle are more likely to use traditional contraceptives over no method and modern methods. 204 THE WORLD BANK ECONOMIC RlVIFW. VOL. 10. NO. I Table 2. Probability of a Woman's Currently Using a Modern Contraceptive Method: Role of Education and Income Women s Including Woma'ss education By womnan s age communitv education and bouse- Tess than 3 S or By ivoman's residence cbarac- Variable only hold income I3S older Urban Rural teristics, Woman's education (completed years)" 1 4.432 3.455 5.888 -0.568 12.033 1.981 3.461 (1.13) (0.88) (1.21) (0.08) (1.00) (0.52) (0.89) 2 (complete preschool) 2.201 1.313 1.586 1.152 5.950 0.581 1.615 (0.81) (0.48) (0.45) (0.26) (0.74) (0.22) (0. 59) 3 3.038 3.176 5.147 0.217 1.911 3.211 3.546 11.19) (1.23) (1.54) (0.05) (0.25) (1.27) (1.38) 4 4.666 4.584 4.111 5.616 7.885 3.112 5.355 (1.85) (1.80) (1.30) (1.30) (1.18) (1.21) (2.11) 5 4.060 3.928 4.546 2.298 0.578 4.071 4.667 (1.97) (1.86) (1.75) (0.63) (0.09) (2.00) (2.21) 6 4.314 3.764 3.344 3.814 2.838 3.611 4.045 (2.00) (1.68) (1.26) (0.89) (0.47) (1.61) (1.80) 7 (complete primary) 8.153 7.320 6.557 9.116 12.743 4.535 7.620 (4.61) (3.85) (2.90) (2.48) (2.58) (2.31) (3.96) 8 7.272 5.843 4.466 15.511 6.633 5.056 6.487 (2.33) (1.82) (1.27) (1.75) (0.93) (1.45) (2.03) 9 (complete Form 2) 10.863 7.848 8.374 2.301 8.970 7.502 8.152 (4.66) (3.08) (2.88) (0.41) (1.58) (2.55) (3.20) 10 11.768 9.102 9.274 5.238 9.068 9.784 8.840 (3.23) (2.41) (2.30) (0.43) (1.20) (2.17) (2.34) 11 (complete Form 4) 13.454 8.902 8.811 14.928 10.776 8.584 8.776 (6.08) (3.50) (3.12) (1.96) (1.93) (2.76) (3.45) 12 or more 21.366 15.895 14.213 29.167 16.010 31.290 16.733 (5.39) (3.79) (2.99) (2.89) (2.21) (3.64) (4.01) Husband's education (comnpleted years)b Completepreschool 1.137 0.516 4.144 -2.313 2.150 0.689 (0.31) (0 .1) (0.64) (0.18) (0.61) (0. 19) Some primary school 1.935 2.633 0.190 3.377 1.549 1.658 (1.10) (1.17) (0.07) (0.68) (0.90) (0.95) Complete primary school 1.729 2.093 -0.805 1.872 1.434 1.451 (0.99) (1.00) (0.24) (0.42) (0.81) (0.84) Complete Form 2 3.314 2.395 4.865 6.580 0.634 2.741 (1.57) (0.97) (1.16) (1.36) (0.27) (1.31) More than Form 2 5.892 6.629 -2.616 8.099 4.293 5.602 (2.76) (2.75) (0.50) (1.71) (1.75) (2.64) ing is actually smaller (albeit not significantly) than for those with one less year of education: surely, they are not learning not to value contraceptives in their first year at secondary school.9 Additional years of secondary school are 9. Many estimates of the relationship between education and wages also indicate steps associated with the completion of particular levels of education; these steps have often been attributed to credentialism (although selection seems like a plausible alternative hypothesis). For contraceptive usage, however, it is far from clear how or why credentialism should play any role. Thomnas and MAaluccio 205 Table 2. (continued) WoVoratn 's InlUheding Wornan s education By 1 ldmian s igi io,iflflUiiity education and bouse- Less that, 15 or By iomnan' s residence cbarac-- Variable only hold incomne 35 older Urban Rural teristics Household assets" Motorcycle or bicycle 3.486 3.0 72 3.328 6.005 2.428 2.8(70 (2.98) (2.23) (1.46) (2.34) 11.89) 12.4 5 Car -0.779 -2.41 2 4.236 -3.280 3.325 -0.741 (0.42) (1.10) (1.09) (1.10) (1.18) (0.40) Radio 0.796 0.716 ().583 0.256 ().792 0.605 (0.67) (0.533 (0.23) (0.09) (0.63) (D.51) Televisioin 4.519 1.92 13.585 5.633 5.042 . 1 89 (2.40), (10.9z)) (3.3 2) (2.02) (1.29) (2 . 78 ) Cattle -2.131 -3.016 1.173 -2.336 -1.514 -1.568 (1.50) (1.82) (0.42) (0.27) 1.16) (1.08) Goats or sheep -3.793 -3.926 -2.834 -13.651) -2.55() -1.684 (2.71) (2.38) (1.0(6) (1.45) (2.0)1) (1.16) Rural residence' -5.305 -0.565 0.222 -3.7 11 0.868 (4.48) (0.38) (0.D3) (1.13) (0.40) Goodness of fit X2(owneducation) 61.41 25.25 1-7.1; 17.69 14.35 22.42 26.43 0.001 [().01] 10.141 10.131 10.281 10.031 n0.01) x2(all covariates) 562.53 596.93 490.50 123.45 261.35 321.35 55.1.08 10.00° [0.001 0.()() 1(.00) [0.0)( 1().(] 1()001 Likelihood -2,087 -2,0)64 -1,51i - 54 -721 -1,323 -2,0)31 Pseudo R2 0.15 0.16 ().18 (0.11 ).2( 0).14 ).17 Numberofobservations 4,201 4,201 3,129 1,()72 1,407 2,794 4_2(11 Note: The values are maximum likelihood probit estiniates transforimed into derivatives. Asymptotic t-statistics are reported in parentheses, and p-values are in brackets. Regressions also inclide controls for age (represented by dummies for each five-year age groLIp), ethnlicity, narital status, and presence of husband in hotisehold. a. See table 3 for a list oii conmmunlily ciaracteristics. b. Dummy variable: Value is I if conditionl is true: ) otherwise. Source: Authors' calculations based On iso and IR i/Macro International (1989, 1991) and ZNYI'( (1992). associated with slightly higher probabilities of adopting modern methods, although the next dramatic step is among women who go beyond "O" levels (twelve years or more): they are 20 percent more likely to he using a modern method relative to women without any schooling. This pattern is remark- ably similar to the relationship between education and fertility, apart from the spike at completion of primary school. Ever-married women are significantly more likely than women who have never been married to be using a modern method. Their probability of adopting contraception tends to rise with age until around thirty and then declines. Women in rural areas are significantly less likely to be using a modern method. 206 THF WORID RANK FCONOMli: RFVIFW, Vol . \N). I In the second column of table 2, husband's education and dummies for owner- ship of assets are added to the regression in an attempt to control for income. Apparently the entire rural-urban gap in contraceptive usage can be attributed to income differences (as measured here): the rural effect becomes very small and is not significant. Recall that in the case of fertility the rural-urban gap persists after controlling for income, although it is much smaller for younger women. Both of these results probably reflect the higher investment in ilfrastructure in rural areas in recent years (since independence). Husband's education is posi- tively and significantly associated with the probability that a woman is a contra- ceptive adopter, hut the effect is far from smooth and is significant only at the top of the education distribution. Income, as measured by the dummies for asset ownership, should not be con- taminated by current labor supply and leisure choices (although they may re- flect previous labor supply). Women in households with a bicycle or motorcycle or a television are significantly more likely to be using a modern method. In contrast, however, women in households witlh goats and sheep are less likely to use contraceptives. The exact reason is far from clear: it may be that this is a poor measure of income (because, for example, no information is provided about the number, type, or value of the animals); it may reflect more traditional atti- tudes among these women; or, it may reflect greater demand for child labor in households with livestock.'( It is most unfortunate that until DHS data on in- come or household resources are collected, it is not going to be possible to ad- dress these issues with these otherwise extremiiely rich surveys. Income (as measured here) does account for part, but certainly not all, of the impact of female education on contraceptive usage. The role of income tends to be greater at the top of the educationi distribution where the estimated educa- tion effects are reduced by about a third; at the bottom of the education distri- bution, the effect is reduced by about 10 percent. While the general shape of the relationship remains intact, the steps at seven and twelve or more years of school- ing are, if anythlillg, even more pronounced. Women are stratified into two age groups in the third and fourth columnns of table 2. The effect of education on contraceptive usage is substantially flatter for younger (fifteen to thirty-five years old) relative to older (thirty-five to forty- nine years old) womel. For example, older women who have comiipleted eleven years of schooling are more than twice as likely to use contraceptives as those who had one less year of schooling, and only half as likely to be adopters than women with more schooling. In comparison, amonig younger women the differences among those who attend secondary school are small except for those IO. Si ice ownership of goats and sheep is uLnrelated to fertility (table Ii, endogeneity is probably niot the dominianit explanation. Moreover, dropping the covairiate hals little effect (oi any 1other covariares other thani cattle ownership (which becomiles absolutely larger and signlificanit). This is because the two covariates are highly correlated (0.621. 0wniership of caLttle, goats, anid shcep is not exclusively a rulal phenNS There can be little doubt that the availability and quality of family planning and health services in the community are associated with higher rates of adop- tion of modern contraceptives in Zimbabwe and these effects tend to be larger for less educated womenl. Public health policy has sought to bring CBDs to all communities in Zhimbabwe: according to our results, this is likely to be asso- ciated with increased adoption of modern methods. Mobile family planning clinics seem to have an even more powerful impact on adoption of contraceptives as does the presence of a general hospital in the area. It is especially important that the impact of these two investments in infrastructure is larger among women with little educationl. Not only do these services have a significant impact on usage but the magni- tude is substantial. For example, if a hospital, mobile family planning unit, and CBD were all introduced to a community where they did not previously exist, then our estimates suggest that contraceptive usage would be raised from 30 percent to 40 percent, on average. This is a very large effect. For example, in terms of the impact on contraceptive adoption rates, it is equivalent to giving a woman who has only completed preschool an additional seven years of school- ing (so she completes Form 2). Education also has a direct, powerful impact on contraceptive usage and fertil- ity, especially amonig younger and urban women for the case of fertility. Part of Thomas and Maluccio 217 the impact of education may be attributed to the role of income and part to unobservable differences among women who complete particular grades (at least for contraceptive usage). It behooves us to better understand the mechanisms that underlie these correlations: one possibility is that education's impact on contraception use and fertility reflects increased labor market opportunities for educated women in post-independence Zimbabwe. 218 THE WORL D BANK ECONOMIC REVIEW, VOL. 10, NO. I Table A-1. Sample Means: Individual and Household Characteristics By woman's age Less than 35 or By woman's residence Variable All women 35 older Urban Rural Women's contraceptive use and knowledge (percent) Know modern method 95.38 94.95 96.64 97.23 94.45 CurrentiV use modern method 27.21 27.96 25.00 33.62 23.98 Currently use pill 23.54 2S.95 16.51 27.65 21.47 Currently use traditional method 4.98 3.80 8.40 2.63 6.16 Ever used modern method 48.42 46.47 54.10 55.86 44.67 Fertility Number of children ever born 2.95 1.86 6.15 2.28 3.29 Number born in last 5 years 0.79 0.83 0.71 0.64 0.88 Completed fertility (women age 45 and older) 6.83 n.a. n.a. 4.99 7.51 Number of births in last year 0.17 0.19 0.11 0.13 0.19 Age at first birth (years) 18.84 18.49 19.49 19.14 18.70 Child survival rate (percent) 91.96 93.30 89.42 94.77 90.65 Hlousehold characteristics Women with completed years of educationi (percent) 0 13.64 10.07 24.07 6.47 17.25 1 1.64 1.28 2.71 0.71 2.11 2 (complete preschool) 4.05 2.81 7.65 1.99 5.08 3 4.86 3.29 9.42 2.63 5.98 4 4.95 4.06 7.56 3.84 5.51 5 9.47 7.96 13.90 5.05 11.70 6 9.19 9.46 8.40 6.18 10.70 7 (complete primary) 21.59 23.62 15.67 22.17 21.30 8 4.71 5.91 1.21 4.90 4.62 9 (complete Form 2) 8.76 10.13 4.76 13.43 6.41 10 3.14 3.99 0.65 4.62 2.40 11 (complete Form 4) 12.26 15.47 2.89 23.67 6.51 12 or more 1.74 1.95 1.12 4.34 0.43 More than primary school 30.61 37.46 10.63 50.96 20.37 Husbands present (percent) 44.58 39.02 60.82 49.18 42.27 Husbands with completed level of education (percent) Preschool 1.81 1.47 2.80 0.64 2.40 Some primary schlool 16.14 10.87 31.53 8.60 19.94 Complete primary school 18.76 17.90 21.27 15.42 20.44 Complete Form 2 9.16 9.24 8.96 12.01 7.73 More than Form 2 13.07 14.51 8.86 23.10 8.02 Age of respondent (yearsl 27.82 23.28 41.08 27.22 28.12 Thomas and Maluccio 2 1 9 Table A-1. (conitinuiied) By uo man's age Less than 35 or By w'o,nan 's residence Variable All Uwomen 35 older Urban Rural Households that own assets (percent) Bicycle or motorcycle 26.66 2 5.95 28.73 26.72 26.63 Car 11.69 12.08 10.54 25.66 4.65 Radio 45.82 48.19 .38.90 74.91 31.17 Television 14.76 15.S() 12.59 39.30 2.40 Cattle 40.66 38.99 45.52 5.12 58.55 Goats or sheep 41.61 40.17 45.80 4.34 60.38 Percentage of women who are: Shona, 78.15 77.82 79.10 69.72 82.39 Married 62.91 56.09 82.84 55.37 66.71 Divorced 10.12 8.15 15.86 12.22 9.06 Resident in rural sector 66.51 64.88 71.27 0.00 100.00 in.a. Not applicable. a. Ethnic group in Zimbalbwe. Source: Authors' calcuL Iationis biased on ( [i aind IRIi/Maicro Intern a tionad (1989). 220 IHi( WORLD BANK ECONO).MI RHVII Wi. VOL. 10, NO. I Table A-2. Sample Means: Community Characteristics Cluster Sample women Variable level All Urban Rural General infrastructure (per-ent) General hospital 50.1 51.1 98.2 27.4 Rural hospital 25.7 26.3 7.7 35.7 With electricity 18.6 19.1 7.7 24.8 Clinic built since 1980 45.5 47.4 24.4 59.1 Clinic with electricitv 47.3 45.5 98.1 19.1 Visited bv health worker 28.7 62.9 38.8 75.1 Visited by mobile nimunization unit 42.5 43.6 0.0 65.6 Visited hv mobile family planining unit 61.7 30.7 9.9 41.2 Communitv-based distributor (percent) Visited hy (CB) 60.5 62. 8 34.5 77.1 CBD has sample kit 50.1 51.3 31.4 61.3 cBn has bike 39.5 39.1 20.3 48.5 (:BI) has hlood pressure gauge 20.8 21. 6 15.5 24.7 CBD has stethoscope 19.7 20.5 13.9 23.8 (:B) has taken ZNFPN course 50.8 51.4 34.5 60.0 Clinic characteristics Number of needles in stock (thousands) .3.0 2.7 2.2 2.9 Lap-kits (percent) 3.6 2.7 1.6 3.2 Distribute condoms (percent) 92.2 91.1 98.4 87.4 Distribute Depo Provera (percent) 15.0 15.9 21.7 13.0 Distribute other methods (percent) 41.3 40.3 67.8 26.5 Suggest natural methods (percent) 6.6 6.1 5.4 6.5 Have at least one doctor (percent) 13.2 10.1 25.9 2.1 Number of nurses 8.( 7.2 14.9 3.3 Sample size 166 4,201 1,407 2,794 Source: Authors' calculations hased onc0so and RDu/M'acro International (1989). Measures ofgeneral health infrastructure are drawn from csO and IRD/Macro Iiternational (1991 ): all other measures are drawn from ZNi'(: ( 1992). Tbomas and Maluccio 22 1 REIERFNCES The word "processed" describes informally reproduced works that may not be com- monly available through library systems. Adamchak, Donald J., and Michael T. Mbizvo. 1990. "The Relationship between Fertility and Contraceptive Prevalence in Zimbabwe." Ittternational Family Planning Per- spectives 16(3):103-06. Becker, Gary S. 1 98 1. A Treatise on the Family. Cambridge, Mass.: Harvard University Press. Boohene, Esther, and Thomas l)ow. 1987. "Contraceptive Prevalence and Family Plan- ning Program Effort in Zimbahbwe." International Famnily Planning Perspectives 13(I ): 1-7. Cochrane, Susan 1983. "Effects of Education and LJrbanization on Fertility." In Rodolfo A. Bulatao and Ronald 1). Lee, eds., Determinants of Fertility in Developing Coun- tries. New York: Academic Press. Cochrane, Susan, and Samnir Farid. 1990. "Socioeconiomic Differentials in Fertility and Their Explanation." In George T. F. Acsadi, Gwendolyn Johnsoll-Acsadi, and Rodolfo Bulatao, eds., Population Growth and Reprocisition in .Sub-Saharan Africa: Tech/ni- cal Analyses ot Fertility and Its Consequences. Washington, D.C.: World Bank. (:so (Central Statistical Office) and IRI) (Institute for Resource Development)/Macro International. 1989. ZimnbabvLe Demzographic and Health Survey, 1988. Harare, Zim- babwe, and Columbia, Md. - 1991. Zimbabwe Service Availability Survey, 1989/90. Harare, Zimbabwe, and Columbia, Md. Efron, Bradley. 1 982. "'The Jackknife, the Bootstrap and Other Resampling Plans." SIANt (.BNts-NYI Monograph 38. Philadelplhia: Society for Industrial and Applied Math- ematics. Feyisetan, Bamikale J., anid Martha Ainswortlh. 1996. "Contraceptive Use and the Qual- ity, Price, and Availability of Faamily Planninig in Nigeria." The Worldl Bank Eco- nomnic Review' 10(1): 159-87. Frankenberg, Elizabeth. 1995. "The Effects of Access to Health Care on Infant Mortal- ity in Indoniesia." Health Transition Rev'iew' 5(2):143-63. Gertler, Paul J.,and John Molynleaux. 1993. "How Economic Development and Family Planining Combined ro Reduce Indolnesiani Fertility." Deniography 31(l):33-64. G(uilkey, David, and Susan Cochrane. 1992. "Zimbabwe: Determinants of Contracep- tive UIse at the Leading Edge of Fertility Transition in Sub-Saharan Africa." Univer- sity of North Carolina, Chapel Hill, Dept. of Economics. Processed. Huber, Peter 1. 1967. "The Behavior of Maximumln Likelihood Estimates under Non- Standard Conditions." In Proceedings of the Fifth Berkeley Symposium on Math- ematical Statistics and Probability. Berkeley, Calif.: University of California Press. Mauldin, W. Parker, and John A. Ross. 1991. 'Family Planining Programs: Efforts and Results, 1982-89." Stiutdies in Family Planninig 72(6):3,50-67. National Resear-ch CouLIcil. 1993a. Factors Affecting Contraceptive Use in Sub-Saharan Africa. Washington, [).C.: National Academ-y Press. 1993L. l)emographic Change in Sub-Saharan Africa. Washington, D.C.: Na- tional Academiiy Press. 222 [HF W()RI [) BANK FC(,ONo.\11(: RFVIFW, VOl.. I U. NC. I . 1993c. I)emograpbic Effects of Economic Reversals in Sub-Saharan Africa. Washington, D.C.: National Academy Press. Pitt, Mark M., Mark R. Rosenzweig, and Donna M. Gibbons. 1993. "The Deter- minants and Consequences of the Placement of Government Programs in Indone- sia." The World Bank Economic Review 7(3):319-48. Rosenzweig, Mark, and T. Paul Schultz. 1985. "The Supply of and Demand for Births: Fertility and Its Life Cycle Consequences." American Economic Review 75(5):992- 1015. Rosenzweig, Mark, and Kenneth Wolpin. 1982. "Government Interventions and Household Behavior in a Developing Country: Anticipating the Unanticipated Con- sequences ot Social Programs." Journal of Development Economics 10:209-25. 1986. "Evaluating the Effects of Optimally Distributed Public Programs." Ameri- can Economic Review 76:470-82. Thomas, Dunican, and John Maluccio. 1995. Contraceptive Choice, Fertility, and Pub- lic Policy in ZimbabWe. LSNIS Working Paper 109. Washington, D.C.: World Bank. Thomas, Duncan, John Strauss, and Maria-Helena Henriques. 1990. "Child Survival, Height for Age and Household Characteristics in Brazil." Journal of Development Economics 33:197-234. Way, Ann, Anne R. Cross, and Sushil Kumar. 1987. 'Family Planning in Botswana, Kenya and Zimiibabwe." International Family Planning Perspectives 13(1):7-1 1. Wilkinson, Marilyi. 1992. "Dimensions of Service Delivery and Their Impact on Pill Use in Zimbabwe." Johns Hopkins University, School of Public Health, Baltimore. Processed. ZNFPC (Zimbabwe National Family Planning Council). 1992. "Zimbabwe: A Situation Analysis of the Family Planning Programme." Harare. Processed. Coming in the next issue of THE WORLD BANK ECONOMIC REVIEW May 1996 Volulme 10, Nuimber 2 A symposiulm on stock mnrkets and econom1ic development, incluiding ... * An Overview by Asli Demirgiic-KuLnt and Ross Levine * Equity Markets, Transactions Costs, and Capital Accumulation: An Illustration by Valerie R. Benciveenga, Bruce D. Smithl, and Ross M. Starr * Stock Market Development and Financial Intermediaries: Stylized Facts by Asli Demiirgii4-Kut and Ross Levinie * Stock Market Development and Financing Choices for Firms by Asli Demirgfic-Kunt and Vojislav, Maksimo,)ic * A Measure of Stock Market Integration for Developed and Emerging Markets by Robert A. Korajczyk * Stock Market Development and Long-Run Growth bil Ross Levitie and Sara Zervos -~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~132,5 -~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~