Labor and women's nutrition : a study of energy expenditure, fertility, and nutritional status in Ghana

Policy Research WORKING PAP-ERS Agricultural Policles Agriculture and Rural Development Department The World Bank October 1992 WPS 1 009 Labor and Women's Nutrition A Study of Energy Expenditure, Fertility, and Nutritional Status in Ghana Paul A. Higgins and Harold Alderman Women's nutritional status is reduced greatly by certain kinds of energy-expending work (especially agricultural tasks) and by "maternal depletion syndrome" in women with high fertility. Policy ResearchWozkingPapes disseminate the findings ofwork in progress ant encourage the exchange of ideas among Bank staffand allothers ltC d in developmentissues.Thesepapers. distributedby theResearchAdvisory Staff,cany thenarnes of the authors. reflect onlytheirviews.andshouldbe used and cited accordingly.The findings, interpretations. and eonclusions are the authors'own.Theyshould not be attributed to the World Bank. its Board of Direetors. its management, or any of its member countries.  Policy Research| 1 3 3 _~1N Agricultural Pollels WPS 1009 This paper - a product of the Agricultural Policies Division, Agricultural and Rural Development Department-is part of a largereffort in the department to monitor the impact of agricultural policies on rural poverty. Copies of the paper are available free from the World Bank, 1818 H Street NW, Washington, DC 20433. Please contact Cicely Spooner, room N8-039, extension 32116 (October 1992, 41 paZes). Economic approaches to health and nutrition tive nutritional decline. But the "maternal have focused largely on measures of child depletion syndrome" remains controversial. nutrition and related variables (such as Much of the evidence to date has been impres- birthweight) as indicators of household produc- sionistic - or the results of studies based on tion of nutritional outcomes. But when dealing small, nonrandom cohorts. with adult nutrition, economists have to address an issue that has generated tremendous contro- Higgins and Alderman used a two-step versy in the clinical nutrition literature. instrumental variables technique to get consistent estimates of the structural parameters. Energy That issue is heterogeneity in an individual's expenditure, as embodied in individual time energy expenditures. Preschoolers' energy allocations over the previous seven days, was expenditure also differs, but the differences are found to be an important determinant of small enough to be ignored. Not so for adults, women's nutritional status. Time devoted to whose waking hours are devoted mostly to labor agricultural tasks, in particular, had a strong activities the energy costs of which vary enor- negative effect. mously. Variables measuring time allocation to various types of labor tasks were used to proxy The results also appear to confinn the differences in energy expenditure. existence of a maternal depletion syndrome. Perhaps more important, evidence was found of Parity has also been hypothesized to be an a substarntial downward bias of the calorie- important detenninant of female nutritional elasticity estimate when the energy expenditure health in high fertility countries - with rapid proxies were excluded. reproductive cycling contributing to a cumula- ThePolicy ResearchWorking PaperSeriesdisseminates thefindings of work under way in theBank. Anobjectiveoftheseries is to get these findings out quickly, even if presentations are less than fully polished. The findings, interpretations, and conclusions in these papers do not necessarily represent official Bank policy. Produced by the Policy Research Dissemination Center  LABOR AND WOMEN'S NUThITION: A STUDY OF ENERGY EXPENDITURE, FER LIY, AND NUTRIONAL STATUS IN GEANA- Paul A. Higgins Tulane UniversitY and Harold Alderman Aoricultural Policies Division Agricultural and Rural Development Department World Bank * Project initiated while the authors were, respectively, research support specialist and senior research associate at the Comnell UniversitV Food and Nution Policy Program. Partial support for the project from the Social Dimensions of Adlustment project of the, World Bank, and from U.S. Agency for Intemational Development Cooporotive AgreementAFROOO-A-O045-00, is gratefully scknowledged. The uthors would like to thank Insan Tunall for conmments and suggestions.  TABLE OF CONTENTS Page Introducion ................ I Methodology ................ 4 Data ................ 13 Results ................ 16 Conclusio.s and Discussion ................ 20 References ................ 21 LIST OF TABLES Tablel .... 26 Table2 .... 28 Table3 .... 30 Table4 .... 31 Table5 .... 33 Table6 .... 35 TableAl .... 36 TableA2 .... 38 TableA3 .... 40  1. Introducion |he study of nutrition has been among the more fruitful applications of the economic theory of household production. In addition to using nutrition as an indicator of welfare, economists have incorporated nutritional variables into studies of labor productivity and wages, poverty, health, and ferdlity (Behrman and Deolalikar). Yet there is a danger in such intellectual border crossings; while nutritional status reflects the state of health of an individual as influenced by the intake and use of food or nutrients (Gibson), nutrient intake is only one in a complex of determinants of nutritional status. Diseases and parasites, and the individual's genetic endowment bave come to be recognized as important covariates conditioning the body's utlization of ingested food. But energy expenditure has been virtually ignored in economic approaches to the subject. Nutritional status is largely the result of individual net energy bance (Beaton 1983b). It is determined, in other words, by the person's energy epndkure as well as her calorie intake. Treating malntrition solely as a problem of inadequate food or nutrient availability can lead to perverse results - - for example, in the extreme, food-for-work progrms may fail to improve the nutritional status of participants if the increased labor effort required offsets the effects of the additional food. At the least, it can paint too narrow a picture of the nutrition problem for planners, causing them to disregard the interactions of non-food policies with nutrition, or to overlook possible aveues for improvement, such as the development of labor-saving devices. From a statstical standpoint, of course, neglecting energy expenditure differences in a population is likely to intrduce statistical bia. Many studies of nuritional status have reported sang positve income effects in reduced form or hybrid nutrtion equatons, even when input such as food or nutrien, morbidity, and health variables are induded. Since income per se is not an input into nutrition production, it is reasonable to ask whether the magnitude, perhaps even the sgnificance, of thes results may be the result of ms- 1  2 specification. If aveag onergy expenditure per unit of time detcreases with income in a population (or with assets that are correlated with expenditures), the effect of reduced energy expenditure could be incorrectly attributed to rising income;' if leisure is a normal good, then the income effect on leisure demand would simply reinforce the impact of this misspecification. Weak or insignificant impacts of calories on nutritional status have also been reported in the literature (e.g., Alderman and Garcia for nutritional status of children). While other plausible explanations, notably errors in the measurement of calorie intakes, have been put forward to explain similar puzzling results, it could equally well be the result of bias due to specification error. Calorie intakes and requirements ,ro dilectly correlated with levels.of individual energy expenditure (James and Schofield), whUe the energy costs of activities are presumably negatively related to indicators of nutrtional stants such as body weight or adiposity, other things being equal. Thus, failing to account for differences in energy expenditure would tend to bias the coefficient on calories downward, quite apat ftom any memen error effects. 'he issue of energy expenditure is especially important in studies of adult nutrition, since adult enegy use can be expeced to vary systematicaily within a population depending on activity level.2 An individual's energy expendiure is determined by her basal metabolic rate (BMR), and by the energy cm of her daily activities. The former is stochastic, and generally unobserved, but correlated with A, gender, and body ma (James, Ferro-Luzzi, and Waterlow). On the other hand, the non toc component of energy expediu is a function of the individual's time uses, and the intensity wih wh she puru them nam nd Schofield). 'Bonis and Haddad discuss this possibility with respect to the estimation of calorLe demand. 2Thls in not to may that energy expenditure can always be safely ignored in child studies (see Beaton 1983b).  3 Note that this problem is relat, though not identical, to the question of heterogeneous nutriet requirements across individuals (and over time) which currently bedevils the clinical nutrition literature (see B&tn 1983a, b; also Dasgupta and Ray). Whether individual requirements are fixed or adaptive, however, they are well-predicted by the energy costs of current activity levels, along with the aforementioned covariates of BMR (James, Ferro-Luzzi, and Waterlow). Hence, including current engy expendiure, or some indicator of time use correlated with energy expenditure, along with age, height and gender seems a way of minimizing this source of bias in a nutrition regression, providing, of course. that the simultaneity of these choices can be appropriately modeled. This study examines the determinants of the nutritional stas of adult women, using household survey data from Ghana. Its main contribution lies in exploiting time use data to estimate the contribution of individual energy expenditure differentials in determining nutritional status. The role of energy expenditures in contributing to female malnutrition is potendally mom important in sub-Saharan Africa than anywhere else in the world. African women tend to spend a ratively higher proportion of their time performing physically demnanding tasks, with relatively less leisure time, due mainly to their central role in agricultural production and distribution, and a lack of labor saving devices (Mueller; Lawrence et al. 1985; Singh et al.; Lamba and Tucker; Mebrahtu). Marked seasonal swings in energy expenditure, as well as in body weight and composition, and food availability, have also been documented among African women, especially in nral areas (Lawrence t al. 1989; Reardon and Maton). A secondary focus of the analysis is the role of fertility. In general, weight increases wih pari. Among undernourished, bigh fertility populations, however, indicators of nutritional status based on weight may decline with increased parity (Adair). Sub-Saharan Africa's average fertility rate (6.5 per woman, compared to 2.7 for East Asia and the Pacific, 2.0 for South Asia, 3.3 for Latin America and  4 the Caribbean3) Is the highest in the world. While acceptance of the notion of a maternal depletion syn. -rme is not universal (Winikoff and Castle express skepticism, for example), recent empirical evidence from a variety of settings suggests that rapid reproductive cycling indeed contributes to maternal nutritional depletion in high fertility coui tries (Merchant and Martorell; Huffman, Wolff, and Lowell; Adair et al.; Merchant, Martorell, and Haas 1990a, b). Since none of these studies treated parity as an endogenous choice variable, however, the possibility of statistical biased results cannot be ruled out. 2. Methodology 2.1. Theoretical Model. The theoretical model used here is based on those employed by Rosenzweig and Schultz (1983) and Schultz. Household members are assumed to behave as joint welfare maximizers with respect to individual health status, consumption, and time allocation.' Hovsehold utility is derived from both purchased and home produced goods, including nutrition and health. A joint household preference function governing household decisions over this choice set takes the frrm: i(1) U = U(Hf, Ci, C', LI', I, A), i =l,..J, where H' is the health status of household member i, C,: is member i's nonfood consumption vector, Cf is i's food consumption vector, L' is i's leisure time, gs is an unobserved variable capturing tastes and norms, assumed to be exogenous to current consumption decisions, and I Is household size. 3According to the World Development Report 1992. 4This assumption has been criticized for ignoring bargaining between household members (e.g., Folbre (19861; Manser and Brown; McElroy and Horney). Different intra-household models can be tested using individual specific non- labor income. The distinction between the two model. is often small in empirical applications, however (cf. exchange between Folbre (119841 and Rosenzweig and Schultz (19841), s0 Occam's Razor would seem to favor the former.  Tho relationship between the nutritional status of each household member and nutrient and health inputs (as conditioned by the individual's health endowment and the household and community health environment and infrastructure) is governed by a production function of the form: (2) H' H(N', A', B', T', FP; Di, S, M, 0i), i= ,...,I where Ni, Ai, B', and E are vectors of member i's recent nutrient intakes, morbidity episodes,5 use of health care services, and energy expenditure, respectively; P' is i's total parity; D' is a vector of other fixed, observable individual characeistes of member i affecting her nutritional status; S and M are vectors of household and community fixed factors, respectively, that affect th6 nutritional status of household members; and VI is i's (unobserved) health endowment. The household maximizes (1) subject to (2) and its full incame budget constraint, generating input demands that enter the right hand side of (2) and which take the general form: (3) Z F r(Y, P; D, s, M, O,i= 1,..., where Z is a placeholder for (Ni, A', Bi, Ei, F), Y is exogenous income, and P is the complete vector of prices, broadly defined to include time as well as money costs. Note that in this specification neither prices nor income etr direcy int the production of health. Instead, they affect (2) idirectly, via the demand for inputs. 2.2. Indviduol Heterogenei. 5Because diseases can reduce the absorption of nutrients consumed, as well as depressing the anpptite, while fevers raise metabolic rates, the morbidity indicator may be thought of as conditioning the nutrient intake variable.  6 Empirical applications of the above modea face several potential pitfalls. Perhaps best known is the problem of unknown individual heterogeneity - in terms of the current specification, the inabiity of researchers to observe ,qi. To illustrate, consider the example of a household in which some members are inherently more robust than others, and thus better able to weather short term shocks such as food shortages. Family members are likely to be aware of this, and in lean periods may choose to allocae relatively more food to those who most need it in order to survive. Typically, researchers cannot observe such differences in individual endowments; yet the observed levels of some health and nutrient inputs will undoubtedly vary according to this individual attribute V. The resulting correlation between the inputs and the error term in equation (2) will bias the coefficients if they a:e estimated using ordinary least squares. There are essentially three possible responses to this difficulty. The first is simply to estimate equation (2) by ordinary least squares, and live with the possibility of bias. A case can be made for doing so, since the bias may be small, and the remaining options, while consistent, are often relatively inefficient (see Buse).' Another option is to use fill information methods to estimate the production function and input demands as a simultaneous system. This was the approach chosen by Rosenzweig and Schultz (1983), for example, in their analysis of birth weight production with endogenous inputs in the United States. Guilkey et al. also used this method in their study of birth outcomes in the Philippines. Explicitly modeling the full structural system is appealing, but this requires the researcher to specify the complete structure of the model, possibly increasing the likelihood of specification error. Perhaps more to the point in many applications, they can make insupportable demands upon the data set. The remaining option - which is related to the second - is to use an instrumental variables (IV) esimator. This method also may impose heavy demands on the data set. It is often difficult in practce 'For this reason it is interesting to compare the estimates of the preferred model, presented below, with the OLS estimates which appear in an appendix table.  7 . to find idendfyg restrictions for more than one or two Wndognous haldth or nutriet inputs, when many more are usually required. Even when sufficient plausible restrictions are avalable to identify the model, the instruments may perform poorly, leading to esdmates of the strucwal coefficients which are imprecise. Idoally, the vector M In equations (3) should contain a compete set of prices and wages, as well as other community variables affecting the demand for nutrien and halth inputs; examples include roads, distance to nearest clinic, quality of available health care services, clima, prevalnce of disease vectors In the local water supply and environs, local dietary and other customary practices, and the like. Ofto these are not observed; with the excepdon of local market prices and some locational indicators this is unfortunately true of the data set we use here. As is true of many integrated household data sets, however, ours was generated using cluster sampling techniques. Each cluster represents a single market and a relatvly homogeneous group of households, and interviews within each cluster were conducted over a short period of time. As such, there is likely to be virtually no intra-cluster price variation, while al of the other variables in M are by ddetion constant within a cluster. This fact suggests the possibility of takng the efcts of these missing variables in the instrumenting equations into account using clustr fixed effects or similar techniques. lbls is discussed in the next section. 2.3. Jdeadfyng Me Effects of Mfsing Comnwday F=tors Consider the following esdmatig equons, which may be viewed a linear approximations to equations (2) and (3): (4) Hi = ZpH + DHP + S,NtH + MH8H + e + (5) DZ= Dzz+SZP +J4+ 4  where the v subscript indexes villages (clusters) and i indexes individuals. As before, Z represents the vector of inputs into the production of nutrition; D, S and M denote observable individual, household, and village fuctors, respecively, which affect nutrition production or input demands; the v and X tem are unobserved personal and village characteristics. If most elements of the vector ?# In equation (5) are unavailable, then the unobserved component ,Z may contain the bulk of the information explaining the use of some inputs, meaning that any predicted values derived from the estimated coefficients of equation (5) are likely to he inefficient instruments. If any of the included predetermined variables are cor:elated with coz, then the parameter estimates from (5) would also be biased. A possible solution to this problem would be to use a community fixed effecs model to estimate (5), whereby the data are deviated from their cluster means7: (6) 2Z . (DZ - D)Z + (Sz -SZ)IZ + i Equivalently, this equation can be expressed in terms of dummy variables: (7) 7 = Dz + SY + X + i where X, is a vector of cluster dummy vaiables, such that the jth element is defined to be unity if individual i lives in cluster j, and zero otherwise. As long as all obsertions within a given cluster are collected in the same time period, then equaion (7) Is nearly equivalent to the more richly parameterized equation (5). ?Strauss (1990) uses a similar method to deal with the problem of missLng household variables in a study of child nutrition in Cots d'Ivoire.  9 Thus, the total variance in equation (5) may be partitioned into the within-cluster sum of squares about the cluster means, and the sum of squares of the cluster means about the grand mean, and their res- pective effects treated separately. Because Mz and dZ are constant across households in a given cluster, they are differenced out when the data are deviated from their means and thus need not be considered in the estimation of (6). For this reason, there is no possibility of bias due to the correlation between M! and c.z in these regressions. Thus, ,Sz and -yz may be consistently estimated. If consistent estimates of these pararneters were the primary goal of this exercise, then cluster fixed effects would be fully appropriate. As a means of obtaining the IV estimator of the nutrition production function, on the other hand, using cluster fixed effects to construct instruments may not be the best choice. Note that the vector MZ is dropped when estimating (6); yet MZ is apt to contain many exogenous variables useful for identifying instruments, even though it may be relatively sparse for the reasons given above. Furthermore, the fixed- effects estimator of equation (6) can use only the information contained in the intra-cluster variation; inter-cluster variation is ignored. For these reasons, the fixed-effects estimator of (6) is likely to be imprecise. The cluster means of the observed Z4,., which implicitly contain much of the information on the effects of the unobserved community variables in explaining inter-cluster variations in input use, may be included as instruments in the estimation of equation (4) along with the predicted deviations. The approach taken ifn the present studv is slightly different. Pither than estimating equation (6) in fixed effects, the cluster mean of the LHS variables in equations (5) is included as an instrument in equations (5); in effect, SZzZ y + MZbz + w0 is replaced by Z,.' As discussed above, the community means implicitly contain information on the missing cluster characteristics useful for identifying the instruments. While the pathways by which the missing community variables affect input levels cannot 8A similar technique is used by Alderman and Garcia.  10 be identified using this method, the impact of these missing variables on the production of nutridon, as they operate through the demand for inputs, can be consistently estimated.9 2.4. Model Specification In this study, women's nutritional status is measured by Quetiet's body mass index (hereafter BMI), defined as (weight/height), the most commonly used indicator of nutritional status for non- pregnant, non-lactating adults.'