SWP302 The Economic Theory of the Household and Impact Measurement of Nutrition and Related Health Programis World Bank Staff Working Paper No. 302 October 1978 The views and intef:pretations in this document are those of the authors and should not be dttributed to the World Bank, to its affiliated organizations, or to any individual acting in their behalf. Prepared by: Dov Chernichovsky Development Economics Department Research Project No. 671-38 opyright © 1978 DO NOT he World Bank PUB 818 H Street, N.W. HG /ashington, D.C. 20433, U.S.A. 3881.5 R E M OV E .W 57 W67 no .302 The views and interpretations in this document are those of the authors and should not be attributed to the World Bank, to its affiliated organizations, or to any individual acting in their behalf. WORLD BANK Staff Working Paper No. 302 October 1978 THE ECONOMIC THEORY OF THE HOUSEHOLD AND IMPACT MEASUREMENT OF NUTRITION AND RELATED HEALTH PROGRAMS This paper conveys in broad terms an economist's approach to concepts and measurement of nutrition and related health programs. It is an outgrowth of an interdisciplinary research project involving the School of International Health of the Johns Hopkins University and the Development Economics Department of the World Bank. This project, which summarized the Narangwal intervention experiment in India, and policy interventions in general gave rise to the need for laying out informally the conceptual framework of the economic theory of the household and the relevance of this theory to evaluation research in nutrition. This paper is an attempt to meet this need by outlining, with no loss of generality, how this theory and its statistical adjunct--econome-. trics--can be used for suggesting testable hypotheses concerning program impact, defining outcome measures, and measuring program impact. Prepared by: Dov Chernichovsky Development Economics Department, Research Project No. 671-38 Copyright () 1978 The World BaLnk 1818 H Street, N.W. Washington, D.C. 20433, U.S.A. TABLE OF CONTENTS Page Introduction 1 I. Intervention Programs and Household Economics 2 II. An Economic Formulation of Health-Related Outcomes 7 III. Impact: Measures: Basic Problems and Solutions 10 IV. An Econometric Approach 16 V. Impact Measurement: Econometrics and Social Experiments 23 VI. Conclusion 34 Annex 36 Bibliography 42 LIST OF TABLES Table 1. Regression Coefficients, Carbohydrates (in grams) and 18 Vitamin A (in int'l units) as Dependent Variables Table 2. Regression Coefficients, Child's Weight in Kg. as a 20 Dependent Variable, Ordinary Least Squares (OLS) and Two-Stage Least Squares (TSLS) Estimates Table 3. RLegression Coefficients, Child's Weight in Kg. as a 25 Dependent Variable THE ECONOMIC THEORY OF THE HOUSEHOLD AND IMPACT MEASUREMENT OF NUTRITION AND RELATED HEALTH PROGRAMS Dov Chernichovsky* World Bank INTRODUCTION Household economics--as a theory of choice--offers a conceptual framework in which to investigate the family's responses to changes in its environment. Since such changes can be induced, for example, by nutrition and health intervention programs, this framework can be useful for policymakers and planners in formulating hypotheses about the effects of such pro- grams. Econometrics, the complementary statistical extension of economic theory, furnishes a versatile statistical framework for testing these hypotheses and quantifying the effects of such programs, as well as in- creasing our basic knowledge about the interactions between the programs and their social environments. This paper conveys--in broad terms--an economist's approach to concepts and measurement of nutrition and related health programs. It emphasizes the close link between economics as a behavioral science and the measurement of intervention programs. In the first section, the basic working assumptions and framework of household economics are introduced and related to the concept of an intervention program. This relationship serves to highlight the economist's conceptual point of departure in analyzing */ My colleagues A.A. Kielmann, M.D., and D. Coate, as well as participants in the PAHO conference for which this paper was prepared, should share in any credit for this presentation. However, neither they nor the World Bank should be held responsible for the views expressed here. - 2 - nutrition and health related interventions and measuring their effects. In the second section, the household's behavioral objectives that are used to specify and measure the outcomes of intervention programs are discussed. These objectives provide the conceptual framework for considering various properties of outcome variables and measurement problems in the third section. The econometric approach to socioeconomic studies in nutrition is discussed in the fourth section. It serves as a basis for discussing in the fifth section the basic differences between an experimental approach and an econometric approach in measuring program impact. I. INTERVENTION PROGRAMS AND HOUSEHOLD ECONOMICS Health-related environmental programs like malaria eradication and largely mandatory programs like smallpox vaccination apparently have been relatively successful in meeting their objectives (World Bank 1975: 38-39), while many personal health and nutrition programs appear to be less success- ful. A common feature of the environmental programs and the mandatory programs is that their implementation does not require individuals or households to choose how to respond. The environmental programs do not deal directly with individuals or households, and mass vaccinations may leave little or no room for individual or household choice. The apparent failure to detect success in personal health and nutrition programs can be attri- buted to many factors. It can result from the lack of data and statistical tools to measure success. However, it can also result from the target populations' particular response, or lack of response, to the programs. - 3 - Malnutrition results largely from a combination of individual and household consumption behavior and hygienic decisions and practices. This fact limits the feasibility, economic and otherwise, of mandatory and effectively controlled nutrition programs; that is, these programs are most likely to leave individuals and households choices of (a) whether to parti- cipate in a program and (b) how to use resources that become available through it. It is the existence of these choices that makes the economic theory of the household a useful evaluation tool, the importance of which corresponds to the potential degree of voluntarism in individual and house- hold responses to a program. The household, whether a nuclear or an extended family, is the basic socioeconomic unit that makes most decisions about investment in human beings and about consumption. 1/ The significance of a household and an individual to the community is not limited to their role as components of a sum. An individual's education, and particularly his health, often affect the well-being of others in the community; communicable diseases exemplify the interdependence between an individual and his community. This interdependence and certain moral norms concerning the distribution of well-being among households in the community provide much of the basic rationale for health-related intervention programs. Economics usually views the household as a harmonious microcosm that makes deliberate and rational decisions. This is a basic working assumption employed to identify the systematic part of human behavior by using conceptual parameters and measurable variables. This assumption is eventually 1/ It is important to realize, however, that in some traditional cultures, tribal o;r village governing entities might make important decisions about investment in human capital. - 4 - "modified" in econometric analyses by taking into account unsystematic var- iations of behavior. The economic analysis of household behavior can be summarized as follows. 2/ Households and individuals engage in activities to produce "ends," or consumption commodities, that render utility. 3/ These ends compete for the household's scarce human and nonhuman resources because producing more of one commodity implies producing less of others. A change in the household's environment can change (a) the household's income or wealth, which determines how much a household can produce, (b) the commo- dities' relative costs or prices, which determine the relative attractive- ness of different commodities, and (c) the household's tastes and prefer- ences structure. Modifications in household behavior are derived from changes, which an intervention program can promote, in one or more of these three. An economic conceptualization of the household's decision making process is sketched in a simplified manner in Figure 1. From the viewpoint of household economics, an intervention program has at least one inherent problem: such a program "disputes" the household's ability and even willingness to realize the social consequences of its choice, and to meet some specified social objectives. This is equivalent in many instances to questioning the rationality and social adequacy of the household's decision-making process and objectives. There are two critical 2/ The approach presented here, in broad terms, is based on traditional demand theory and some extensions of this theory by Becker (1965) and Lancaster (1966). Differences in approach, which bears on the conceptual framework, do (and should) not affect more practical measurement issues. 3/ Those commodities can be abstract as are, for example, "good health" and "services from children." FIGURE 1: AN ECONOMIC CONCEPTUAI IZATION OF THE HOUSEHOLD'S DECISDON MAKING PROCESS Period One (T1) Period Two (T2) Household j Household Endowments Determinants of Choice Choices (Outcomes) I Endowments Time Market " ntervention * _ Time . esur Prices |/ Ti Time~~~~~~~~~~~~~~~~~~~~~~~~~~Tm (Mandatory) T2 \ I ~~Market Heal/tIh Wage ~~~~~~~~~~~~Consumption and and ~~~~Education / / Work Ca/pRi\ap Humans Income HumAss 1 ~~~~~~~Preferences Ass Ts S'cl a-k - 194C0 6 - 6 - implications of this inherent problem. First, an effective demand for, or utilization of, program services by the target population cannot be guaran- teed; second, even when adequate demand exists, it may stem from private objectives that are not congruent with program or social objectives. Therefore, a basic requirement for evaluating program impact is to identify the "non-program" parameters that determine the household's demand for program services. This identification should help to indicate how much a household will use the program--when it will do so--and what use it may make of the program's resources. Program utilization is a function of (a) the degree the program serves the household's objectives, (b) the degree it draws on household resources, and (c) the relative attractiveness of other--subjectively conceived--substitutes for program services that meet the household's objectives. These points can be illustrated by common examples. Preventive medical care programs, particularly nutrition programs, are often hard to implement because the target populations may not recognize their usefulness. In many situations, health services may compete with traditional practices that households perceive as substitutes. Program services may not be used, even when people recognize their usefulness, because of the relatively lower costs, or higher benefits, of presumed or proven substitutes, and because of the burden the program may impose on a household's immediate welfare, as when people cannot afford the time and transportation costs to go to a "free" clinic. Program utilization need not be consistent, however, with program objectives. A target household has the capacity to reallocate program - 7 - resources to reach its own rather than the program's objectives. For example, a mother may feed a particular child less than she feeds other family members because that child participates in a school feeding program. Thus, the mother may attenuate and offset the program's specific objectives and may "spread" its impact. Furthermore, through the effect on one aspect of household life, say child mortality, the program may have an impact on other aspects, say fertility behavior. Although these other aspects may not be among the original objectives of a particular program, nonetheless, one should consider them as possible benefits or costs of the intervention. The economic analysis of the household--by relating program services to household resources and objectives--can help to identify (a) the potential uses of program services, (b) program substitutes, (c) the rela- tive attractiveness of the program, and (d) the extent the program draws on household resources. One can thereby hypothesize who program users will be, how they may utilize the program, and subsequently, what the program's impact may be. The formulation of testable hypotheses about program impact aids in identifying variables and relationships that may measure program impact. II. AN ECONOMIC FORMULATION OF HEALTH-RELATED OUTCOMES Program "impact" or "outcome" variables must be identified and discussed in conjunction with households' objectives or ends. This section deals with the household's behavioral objectives, around which we can model family responses to health intervention programs, and on the basis of which we can identify and discuss appropriate outcome measures. - 8 - Economists have long sought to establish adequate measures of welfare, which is taken as the ultimate goal of man's economic activity. In the absence of better measures, monetary income has been used as a second best measure of welfare. Income, which flows from labor and accumulated stocks of assets or capital, serves as a proxy for welfare because higher levels of income mean higher levels of production and thus more commodities that render utility. Measured incomes, however, ignore many utilitarian non- market household activities such as leisure. Furthermore, higher levels of income do not necessarily correlate with greater lifetime well-being when all dimensions of human welfare are considered; the health and nutrition problems of affluent societies are evidence of this last point (Eckholm and Record, 1977). Health status is itself a key element, as well as a good proxy measure of other aspects of human welfare; it can be considered both a consumption item and an investment item (Mushkin, 1962; Grossman, 1972). Good health is "consumption" because it is an end in itself, accounting for a considerable part of that human welfare not measured by income. Health is also an "investment" because it is a key component of human capital that determines the level and duration of one's market and nonmarket activities. As such, it can be linked to some measurable components of earnings and nonmonetary income. This discussion concentrates on the investment aspect of health because this aspect sets the lower limit of potential benefits from health programs. 4/ The discussion is structured around the concept of expected 4/ A rate of return on a health program as an investment will always un- derstate the actual return by excluding the unmeasurable (consumption) utility derived from good health. - 9 - lifetime earnings. In a simplified way, one can define for an individual of age A the present value of his lifetime expected earnings for N years hence- forth by N H -(i E = E Pi. Pie Wi (1 + r) wher ~ H (1+ r)-i where [Pi . Wp i r) ] states the present value of the earnings an individual "expects" at period i. It is given by his (conditional) proba- bility to survive to that period, Pi ; by the probability of his being H physically able to work that year, PH ; and by his average anticipated pro- ductivity--or returns, pyschic and others, from his time-- during that year, Wi. Given a particular time discount rate r, each time-specific expected earnings component is discounted by (1 + r) , which is the present value of s H a unit of earning at a particular future period. 5/ Each term Pi, Pi., or Wi can be regarded as a health outcome; together they encompass the "invest- ment" dimensions in health. 6/ Ps is based on age-specific mortality rates, H i Pi is based on age-specific morbidity rates measurable by lost working days, and Wi can be measured by one's average daily wage rate. The productivity measure or wage rate, Wi, may warrant more atten- tion because, unlike the other terms, it is not usually assumed a direct outcome of nutrition and health. In a given economic setting, defined by the available technology, land, and capital, an individual's wage rate can be taken as a function of his innate physical and mental abilities, and of his physical and; mental capacities acquired through education and work experience (Mincer, 1974.) Since educational achievement and work experience are 5/ W can be regarded as a term net of investment in health and education that affects the levels of all three terms. 6/ For simplicity, we ignore the interdependence among the three terms. - 10 - outcomes of a process that depends partially on health and nutrition, man's productivity, or WV, can be viewed as a function of his health and nutrition, at least in situations of severe malnutrition. Once expected earnings, as specified, are regarded as an indivi- dual's or a family's objective, the economic analysis can derive certain predictions about the household's behavior. The common approach is a maxi- mization procedure by which the household is assumed to enhance Psi , PiH W and other utilitarian ends, subject to various constraints. This procedure sets the tradeoffs among different ends, defines behavioral optima for each, and thereby provides the analytical framework to deal systematically with the relevant aspects of household behavior and to generate a set of refutable hypotheses. 7/ III. IMPACT MEASURES: BASIC PROBLEMS AND SOLUTIONS The specification of the ultimate outcome variable--expected lifetime earnings--and its components is also important for discussing some key statistical issues that relate to impact measurement and to some basic properties of impact measures. Time lags are critical in measuring program impact and present a key statistical problem inherent in intervention programs. The longer the time between the intervention and the measurement of its outcome, the 7/ Anthropological insights into, as well as prior evidence about, the household's view of the underlying investment process are essential for adequate modelling. That is, for example, particular household members may get substantially better diets and "health investment" than others because their (lifetime) earnings are important from the household's viewpoint. Some discrimination in feeding among household members to protect the actual or potential breadwinner is apparent in subsistence settings. - 11 - harder it is to link the program to its hypothesized impact, because one must also account for environmental, biological, and behavioral changes that may also affect the outcome. This issue, which is discussed in more detail in the next section, is frustrating because the impacts of health and nutrition programs take time to manifest and are often spread over individuals' life- times. This problem of time lags warrants both a conceptual and a prac- tical distinction between two types of programs. The first type is a program aimed at enhancing the stock of human capital by an intervention during some critical period of human physical and mental development. Programs involving mothers and children are of this type. The second type is a program aimed at increasing the efficiency of a flow of services from an existing stock of human capital. Programs involving adult workers are of this type (Basta and Churchill, 1.974). The first type of program has long-run outcomes that are often not practical to measure because of the long time lag. Consequently, one must resort to proxies for measurements. The second type has shorter-run objec- tives, primarily increasing productivity and reducing absenteeism, that are immediately observable for adult workers. Termination of a program of the second type should end its effect and thereby provide another means of testing impact. Thus, selecting appropriate outcome measures is more proble- matic for the first, and more common, type of program. Consequently, this discussion focuses on outcome measures for the more general type of program with long-run impact. In evaluating a potential outcome variable, one should consider these questions: - 12 - i. How does the variable relate conceptually to the ultimate outcome, or any component thereof, and to the program under study? ii. How does it relate statistically to the ultimate outcome variable, or any component thereof, and to the program under study? iii. How reliably can it be measured? iv. What complementary data are needed? v. What are the costs of obtaining and using those data? The first question is important because it relates an observed variable to the conceptual framework and to the specific hypotheses to be tested. The other questions bear largely on the statistical aspects of potential data for testing program impact. When all these questions are considered, trade-offs among particular types, or categories, of variables may appear. We can classify outcome variables by three categories: inputs, intermediate outcomes, and ultimate outcomes. The use of program inputs as proxy measures of program impact is common; when, for example, the impact of a school feeding program is estimated by the amount of calories and proteii the program delivers to the target population. This approach has the merit oi being directly related to the program, and also is probably least costly sincE it is integrated with the program. However, it also tends to be the most presumptive since it may depend on hypotheses yet to be tested about relationships between inputs and eventual impact, and it may ignore program- induced behavioral changes beyond program control. This last issue depends critically on the delivery method; an income transfer to the household is easier for the household to divert from program objectives than, say, a directly administered vaccine. - 13 - Intermediate outcomes can be measured by a variety of variables: child morbidity, intellectual development, school achievement, and anthropo- metric measures including birthweight. These measures apply largely to children because theory and evidence suggest that they predict, and thus approximate, eventual health outcomes (Stoch and Smythe, 1976). Although they are not ultimate outcomes, such intermediate variables are "outputs" and in most cases they approximate ultimate outcomes better than program inputs do. Birthweight, for example, predicts relatively well a child's physical growth, at least during the first years and can be used as an outcome measure for maternal care programs; (Lechtig et al., 1975). A child's physical growth and morbidity at an early age may indicate his future mor- bidity, survival probability, and productivity. A child's intellectual development, which is to a degree nutritionally determined, is believed to be manifest eventually in his productivity and wages, primarily through mental development and school achievement (Selowsky, 1976; Taylor and Selowsky, 1973; Grossman and Edwards, 1977.) These intermediate outcome variables raise the problem of lagged program impact since they basically are manifest over time. Hence, linking the intermediate outcomes to the program inputs is even more complicated than using inputs as proxies for outcomes. Complementary data on non- program variables that affect the outcome may become critical for identi- fying the impact of the program. Consequently, collecting data that relate to the intermediate outcome variables requires more elaborate data collec- tion instrum,ents and statistical tools than when program inputs serve to measure outcomes. Ultimate outcome variables involve issues similar to - 14 - those of the intermediate outcomes, but they are more difficult since they pertain to full lifetimes. Outcomes and related variables can be measured on the basis of individuals, households, and communities. The choice of the measurement unit depends on the specific measurement objectives. Policymakers and program administrators are eventually interested in variables that summarize their efforts on a community level. Students of household behavior are also interested in understanding and explaining the distribution of outcomes in the household or the community. Or, they may seek to understand why iden- tical program inputs have a varying impact across individuals and households of different characteristics. While critical for identifying the circumstances under which programs are beneficial, differences in impact may be concealed when we aggregate or, at times, disaggregate data. This possibility must be recog- nized when we define the unit of measurement. One cannot always distinguish household variables from variables pertaining to individuals. At times a variable can be based on a particular household member; at other times, it can be based on a few members. The definition of a household variable based on more than one household member may be complicated because of the low incidence of health related events at the household level. Identification of the target population is a key criterion for the selection and definition of a household variable. If the target group consists of, say, mothers or potential mothers and we are interested in their nutritional status, then the observation is the mother--an individual, as well as the household because in most societies we observe one mother per household. This household then can be described by - 15 - other common household variables such as religion, income, size, or location, etc. The same reasoning applies when the target group consists of children of given age and sex. A problem usually arises when one has to aggregate within house- holds or other units that share the same socioeconomic endowments. This problem appears particularly acute in measuring some intermediate nutri- tional outcomes. For example, when the nutritional status of all school-age children in a household is of interest, one needs to define a summary variable summarizing the nutritional status of these children in that household. The problem is that the number of children as well as age and sex distributions vary across households. This problem can be handled in a few ways; however, while aggregating within the household, one must con- sider the possibility that not all children are treated equally in a given household. The same problem applies to other household variables. For example, two mothers or two family units may live in one household and share a common income. Splitting such a group's income between the two family units, and subsequently treating them as separate units for statistical purposes, may be erroneous if behavioral patterns in extended households differ from the patterns in nuclear households. In most cases when data are treated by averages, a few critical behavioral issues are assumed: that individuals are not discriminated against, and that the behavior of aggregated social units is the sum of the behavior of some other individual units. Before aggregating or disaggregating data, even at the household level, one must see whether these assumptions lead to different predictions. - 16 - IV. AN ECONOMETRIC APPROACH We turn now, before discussing program measurement, to outline some basic features of econometrics that draws on mathematics, statistics, and economic theory to delineate economic relationships empirically. The development of econometrics has been largely influenced by economists' traditional use of non-experimental data to test hypotheses and to verify economic relationships.8/ Economic theory attempts to describe the nature of particular causal relationships, or "structures", by identifying the parameters that make them up, and by indicating the particular functional relationships among those parameters. As indicated previously, these relationships attempt to account, however, only for what is believed to be the systematic part of man's eco- nomic behavior. In reality this behavior also has random elements and may be determined by parameters that we fail either to identify and observe, or to measure accurately. Consequently, the econometric equivalent of the relationships suggested by theory include a "disturbance" or error term added to the "systematic" part of a particular relationship. This term summarizes the random, omitted, and unidentified or inaccurately measured elements of man's behavior. For example, if a particular simple economic relationship is characterized by y = f(X), its econometric equivalent is Yi = f(X i) + vi (i = 1I.... . n observations) where vi is the disturbance or error term. 8/ The term "econometric approach" used here is substantially the same as the "structural equations approach", see Duncan (1975). That is, it is a general application of mathematical statistics not necessarily re- stricted to "economic" variables. The reader familiar with this approach may bypass this section. - 17 - Thus, econometric relationships also are causal relationships and are stated with one or more equations, depending on the underlying structure, each having a single dependent, or "outcome", variable and one or more inde- pendent variables. 9/ Estimating such equations involves obtaining usually, by means of "statistical control", at least unbiased estimates of the parti- cular effect on the outcome variable of each independent variable. 10/ Such estimates are possible when there is no, or only a "small", correlation be- tween the error term and any of the right hand variables, and between any two of these variables. Given a particular economic model, at least three considerations are pertinent in specifying and interpreting econometric relationships: choosing explicit functional relationships, controlling for variables not suggested by theory, and dealing with different and, therefore, competing hypotheses consistent with a particular estimate. The specification of a particular functional relationship can be based on common sense, theoretical and empirical knowledge, and experimentation. For example, we expect caloric consumption normally to level off as income rises, because of saturation and also because of substitution away from calories and carbohy- drates to more "luxurious" proteins. Hence, the specifications of a func- tional relationships between caloric consumption and income should allow for a nonlinear relationship. The two estimated equations shown in Table 1 are a simple example of econometric equations; the effects of certain household variables on 9/ A common introductory textbook to econometrics is by Johnston (1972). See also the Annex, Sections A and B. 10/ The common estimation precedure is regression analysis, which is also a useful descriptive tool. For biological applications of regression analysis, see for example, Snedecor and Cochran (1967:381-418). Table 1: REGRESSION COEFFICIENTS, CARBOHYDRATES (In Grams) AND VITAMIN A (In Int'l Units) AS DEPENDENT VARIABLES (t statistics in parentheses) Dependent Independent Variables Variable North-Urban Region Urbanization Education Adjusted Family Family Elem. School (Nutritient) Income* Size* (Intercept) South** Non-Farm** Rural-Farm**High School** College** R Vitamin S.* 0.1542 0.5130 10.7470 -0.1310 0.0229 -0.0030 0.1852 0.1279 .743 (6.9459) (4.1073) (-4.6454) (0.5313) (-0.0864) (4.4095) (3.7289) Carbohydrates* 0.0085 0.9135 7.8833 0.0579 0.2657 0.1489 0.0486 0.0308 .920 (0.6250) (11.8791) (3.3276) (9.9887) (6.9906) (1.5837) (1.4597) * The logarithms of the variable values were introduced in the estimated equation. ** Denotes the use of a "dummy" variable which takes the value of 1 when an event occurs and the value of 0 otherwise. - 19 - household consumption of vitamin A and carbohydrates are estimated. I1/ These estimates are based on a relationship like that shown as (2) in the Annex, Section B. This relationship allows for the expected nonlinear effect on consumption of nutrients by employing the logarithms of intakes of carbohy- drates and vitamin A, and the logarithm of household income. This functional relationship was also chosen because it allows for a comparison between the sensitivity of the intakes of carbohydrates to income, and the intakes of vitamin A to income, regardless of the different units by which the two nutrients are measured, grams or international units. Such a relationship may impose various restrictions on the estimates; for example, the relation- ship specified presumes no consumption when the household has no current income. Nevertheless, such an unrealistic presumption and other restrictions do not outweigh the advantages of using this particular relationship. Another example of choosing functional relationships involves measurement of child growth by weight and height. Again, nonlinearity concerning age is appropriate here. The estimates shown in Tables 2 and 3 use a quadratic functional relationship that allows for this nonlinearity. These examples show how particular functional relationships are chosen on the basis of prior knowledge as well as practical considerations. Controlling for variables not originally suggested by theory may be a useful me!ans for improving the precision of the estimates, standardization, 11/ This table is drawn from Chernichovsky (1977) where an attempt is made to use economic theory and econometrics to predict and measure the effects of household characteristics on its diet. The coefficients on the logarithmic variables show the percentage change in household consumption of the particular nutrient due to a given percentage change in these variables. The coefficients on the dummy variables show the same but compared with variables which were "left out", or included in the irntercept term. These estimates and others are used here for illustrative purposes only. Table 2: REGRESSION COEFFICIENTS, CHILD'S WEIGHT IN KG AS A DEPENDENT VARIABLE, ORDINARY LEAST SQUARES (OLS) AND TWO-STAGE LEAST SQUARES (TSLS) ESTIMATES (t statistics in parentheses) Type of "Biological" Variables Economic Variables: Occupation Diet Equation Estimation Intercept Sex- Land Agri. Civil No. __________ ______Ae .. Ag e ' Male* Age.Sex (Age.Sex)2 Owner* Labor* Servant* Calories R2 1 (OLS) -0.63771 0.609 -0.008 7.533 -0.563 0.011 0.860 0.038 0.745 0.61 (6.628) (-4.736) (5.526) (-4.280) (3.708) (2.559) (0.110) (1.851) 2 (OLS) 3.173 0.245 -0.019 0.708 0.001 0.55 (3.559) (-1.261) (2.994) (4.125) 3 (TSLS) 1.788 0.112 -0.0001 0.169 0.004 1 (1.246) (-0.077) (0.272) (3.235) C* *Dummy Variables. - 21 - and adding irLformation. 12/ An economic model of behavior may ignore bio- logical factors by assuming them constant. For example, a model that deals with parents' choices of their children's diets and, therefore, child growth (Chernichovsky and Coate, 1977) might assume that the analysis is confined to a hypothetical age and sex group and, therefore, disregard the age and sex variables in explaining variations in child growth. However, for an econometric analysis based on a relatively small sample of children of both sexes and across age groups, one should control for age and sex as illustrated :Ln Table 2, Equation 2. Otherwise, the relationship between child weight and age will be entirely approximated by the correlation between diet and weight across children of different age groups. 13/ This means of control is an alternative to various standardization procedures, like weight for age, used by nutritionists. It is also informative because it depicts growth curves directly across age and sex groups--controlling for other effects--and can be based on relatively small samples. The problem of dealing with competing hypotheses is major. The first and basic issue to address is whether the line of causality implied by a particular relationship is correct. For example, in Table 2, Equations 2 and 3, the effect on children's weight of caloric intakes is estimated. However, the case can be made that the causality also operates the other way: 12/ Note that inclusion of omitted variables, which are not perfectly corre- lated wLth other independent variables, will increase the multiple cor- relation coefficient, or "explained variance" of the dependent variable. This increase should not be a goal by itself. 13/ This procedure can be elaborated to account for interaction between the age or sex variables and others variables; that is, when, for example, a given diet has a different effect on children of different age and sex. - 22 - heavier children consume more calories ceteris paribus. In a case like this it is reasonable to assume a structure where child weight and caloric consump- tion are co-determined. Therefore, one must use appropriate estimation techniques to allow for this codetermination or simultaneity. The estimates reported in these equations exemplify a case like this. The estimated effect of caloric consumption on weight in the third equation is based on a pro- cedure that accounts also for the effect of weight on caloric consumption. This estimated effect differs from the estimate in the second equation where the estimation procedure does not account for simultaneity. 14/ The second issue is to ascertain that the variables indeed measure (or approximate) what they are meant to measure. In the context of nutrition interventions, a good example is the behavioral change in food consumption of families under observation due to presence of an interviewer. Thus, the variable that was supposed to measure the effect of program inputs measures instead another behavioral change induced by the program, thereby confounding the measurement of the program impact. Similarly, the estimated effects of the sex variable on children's nutritional status may measure parents' be- havioral discrimation between sons and daughters, as well as genetic dif- ferences between sexes. In a case like this, it may be impossible to 14/ Two or more variables are simultaneously determined when they are outcomes or endogenous variables belonging to a structure where one of these variables affects the other in some relationships and vice versa in others. Failing to account for this codetermination while estimating one particular structural relationship, may result in a "simultaneity bias" which means that the estimated effect of one endogenous variable on the other way be under- or overstated. On how to deal statistically with such cases, see Johnston (1972: 341-423); for specific application in a nutrition study from which this table is drawn, see Chernichovsky and Kielman (1978). - 23 - discriminate statistically between the two competing hypotheses. 15/ Such cases put an added burden on the scientist beyond getting unbiased estimates; he must be careful in interpreting his findings in light of a variety of behavioral and other relationship that can produce particular statistical results. Some cases of competing hypotheses are more predictable and statis- tically manageable than others. These cases can be generally characterized as those in which one has some notion or knowledge about correlations between only two independent variables. For example, the U.S. food stamp program is designed for low income groups. Consequently, a variable representing household participation in the program is expected to be negatively and highly correlated with the income variable in an analysis of this program's impact across income groups. Therefore, statistical discrimination between the effect of the program as opposed to the effect of income on a particular outcome variable may present a significant problem. A correlation between the disturbance term and one or more independent variables may have similar consequences. Econometrics is concerned with the problems briefly described here and other related ones. The solutions offered are geared, however, to exist- ing data ancd, consequently, are largely based on statistical control. V. IMPACT MEASUREMENT: ECONOMETRICS AND SOCIAL EXPERIMENTS Art illustration based on one of the major "experiments" in nutri- tion can be helpful to the more general discussion that follows about the 15/ For a technical discussion on related issues in the context of a nutrition project, see Nasoetion (1975). - 24 - use of econometrics in measurement of program impact and the relationship between econometrics and social experiments. The estimated regression coefficients shown in Table 3 are based on the Narangwal experimental project. 16/ The first equation--which is equivalent to comparing the mean outcomes of an experiment--shows the esti- mated effects on child weight of various interientions, without controlling for other variables that might affect the outcome (see annex, section E.) The second equation shows the same, but with statistical control for age, sex, and socioeconomic status approximated by caste. The first equation shows that nutritional supplementation has a negative and statistically significant effect on child weight, while medical care and a combination of medical care and nutritional supplementation had no effect. Once we control for the other variables, intervention by nutritional supplementation and by nutritional supplementation combined with medical care show a positive and statistically significant effect on child weight. The Narangwal experiment was not a "true" experimental design, but for that matter, it probably could never be. 17/ Moreover, it represents well the constraints of the kind most field programs face. Treatments were 16/ The discussion here is brief and for illustrative purposes only. It does not represent the results of the entire experiment. For a description of the design of the experiment and its results, see Taylor et. al. (1978). 17/ "Experimental design" refers to use of a "control" group and "treat- ment" group (or groups) selected entirely by a random process. In the purest and simplest type of randomization or experimental design, as opposed to an econometric approach, one is not concerned with identi- fying and specifying particular causal and structural relationships and with estimating their various parameters. Full discussion of experi- mental design is beyond the scope of this paper. The discussion here is based on Campbell and Stanley (1966), Rose and Smith (1968), and Bennett and Lumsdaine (1975). Table 3: REGRESSION COEFFICIENTS, CHILD'S WEIGHT IN KG. AS A DEPENDENT VARIABLE (t statistics in parentheses) Socio-Economic "Biological" Variables Status Types oc Intervention Equation 2 Sex Higher Caste Nutrition Medical Nut/ Supp. + No. Intercept Age (Age) Male (JAT)* Supplement* Care* Med. Care* R 1. 7.96897 -1.12432 -0.08715 0.27777 .064 (-2.45703) (-0.18166) (0.69498) 2. 3.18148 0.00967 -0.00001 0.51932 0.52166 0.77071 0.22200 0.76378 .707 (10.81462) (-4.59184) (4.00037) (3.85227) (2.76640) (0.80560) (3.34425) * Dummy variables. - 26 - applied to entire villages because random assignment of the treatments within villages was socially and practically inconceivable. Randomization on the basis of a population of villages was impractical because it would inflate the program to proportions beyond its financial and logistic means._18/ Two issues had to be considered before estimating program impact. First, although similar, the villages may have been different also in aspects other than the treatments. Second, there was scope for behavioral "self selection"; different people could, by choice, avail themselves differentially of the services, and for that matter, could benefit differentially from given services. Indeed, the combination of these factors is evident in the results presented here. The first equation shows merely a negative correlation between child weight and nutritional supplementation. This is consistent with either or a combination of the following hypotheses. First, the village where children received nutritional supplementation could have had children with a lower average weight, possibly because of nutritional reasons as well as the children's age and sex composition. Second, higher caste and better caring mothers, who have heavier children, could have used the nutrition program more, and more efficiently, than others. 19/ Once we control for these possibilities, by adding the other variables to the estimated relation- ship, we get better estimates of the interventions' impact in the short run as measured by child weight. 18/ Almost none of the major other field experiments, like those undertaken by INCAP, meets the requirements of experimental design. The same is true, for that matter, in the field of family planning (Cuca and Pierce, 1977). It appears that researchers attempted, for practical and ethical considerations, to approach a situation where they have "matching groups". 19/ Other possibilities not discussed here are also plausible. - 27 - This example can be useful in outlining a general (and pragmatic) approach as well as specific criteria for assessing statistical requirements for evaluation research. However, we must first agree on the following points: (a) on pure statistical grounds, (randomized) experimental designs based on large samples are superior to any other approach for "netting out" impact; (b) quantitiative evaluation of programs must be, however, socially efficient; and (c) to meet the efficiency requirement, we must always weigh the value of the additional (marginal) gain in statistical evidence and accuracy versus the marginal costs involved. The last point is a general efficiency criterion that may dictate the use of experimental design, econometrics, and a combination of the two. 20/ Returning to the Narangwal example with the above points in mind, we have to address, first, what it represents. And, second, what do the results mean in terms of the relative efficiency of social experiments. On the first question, in so far as the approach taken in Narangwal is representa- tive of most field experiments, it implies what economists call a "revealed preference"': the best that could and, probably, can be done, given objec- tives and constraints. This does not mean that we cannot have better field experiments. It merely puts a question mark on their economic and political feasibility. On the second question, the results obtained by means of statistical control are most likely biased, or lack "Internal Validity", at least because of the nonzero correlations between any two of the right-hand variables. However, these biases do not appear serious, and the results are fairly 20/ The combined use of econometrics and experimentation is emerging in eco- nomic studies. For example, the New Jersey Graduated Work Incentive Experiment", see U.S. Department of HEW (1973). - 28 - "reasonable." 21/ That is, it is highly probable that a full fledged (and costlier) experiment might not improve substantially the results. While this is a testable proposition, which warrants an experiment, the author is willing to conclude that to determine the impacts of the Narangwal project, the mix of "semi experimentation" and a subsequent econometric approach prove sufficient. To outline an econometrician's general approach to data require- ments for impact measurement, let us consider three interventions aimed at increasing caloric consumption: (a) a national food subsidy program to reduce the prices of cereals to all consumers, (b) a national food stamp program for everyone, and (c) a national food stamp program with randomly selected participants. Assume that the three programs will give parti- cipants identical price reductions. 22/ Program (c' can be evaluated simply and unambiguously, just by collecting caloric consumption data, after the fact, if we are content merely to answer two questions: Did the program increase caloric intakes? If so, how much per family on the average? The answers to these questions could be unequivocal, assuming we gather accurate data and we draw a large enough sample. But this approach would tell us virtually nothing about why some families respond to the program more or less than the average or in the case of program failure, why, there should 21/ The results indicate that using the short term outcome, child weight, nutritional supplementation had a measurable effect. Due to this par- ticular intervention child weight in the sample used here increased by .77 kg., or 9.7%, at the sample mean. The highest zero-order correlation between two of the right-hand variables was .34. It should be noted that a more careful econometric treatment of the data presented here, may improve the results. 22/ Other differences between the programs are ignored. For a discussion of the different implications of programs, like (a) and (b), see Reutlinger and Selowsky (1976). - 29 - have been no significant impact. In more general terms, the experimental program (c) lacks "external validity". Moreover, apart from cost considera- tions, the administration of this program poses social and other hazards that may jeopardize the validity of the experiment. Programs like (a) and (b) are more realistic and common than an experimental program like (c). Taking this reality as a constraint, we stress how one can measure impact using an econometrician's perspective. Although programs (a) and (b) cause identical reductions in prices, there is a funda- mental difference between them. Under the first program, households can buy the cereals only at the subsidized price. Under the second, households can exercise choice in buying cereals with or without food stamps, and they can even sell the stamps. 23/ This difference between the programs is also of key statistical significance. For the first program, all we know is that it started at: some particular point in time and existed to a subsequent point. But for the second, additional important information is potentially available: whether or not, and possibly to what extent, a household used the program stamps for buying cereals. To determine impact in the first instance, panel data are necessary. Preprogram, or "baseline" data, should include the variables believed, on theo- retical and other grounds, to affect the outcome. These data serve to esti- mate the basic relationship between, say, caloric consumption, and its non- program determinants. Once this relationship is estimated, we may also be 23/ Although the food stamps are available to all, one must consider the time and trouble of getting them. For example, a mother with five small children, living at a distance from where food stamps are sold, may find it less attractive to obtain those stamps than a mother of five grown children living nearby that selling station, ceteris paribus. See Annex, Section C. - 30 - able to estimate the added program effect. 24/ For example, suppose the food subsidy is introduced in an agricultural economy and reduces prices of cereals. But if farmers' incomes fall during the program period because of bad weather, for example, the farmers may consume less cereals even at the lower prices. Hence, such a fall in incomes may offset any positive effect of the program. Therefore, by accounting for changes in incomes (or even weather conditions), one can estimate program impact, provided no event significantly affecting the program outcome occurs simultaneously with the program in a way to confound our inferences. In the second instance, the national food stamp program (Program B) baseline data may not be needed if we can analyze statistically the variance in households' utilization of the program. This possibility depends on the relationship between program utilization and other variables. At least from a behavioral standpoint, it is almost inconceivable that households of dif- ferent characteristics will use, and even benefit from, the program iden- tically. Therefore, it may be possible to study the determinants of program utilization as well as the interaction between program inputs and household characteristics. That is, what are the effects on utilization of various income and education levels? How do households of different levels of income and education benefit from given program inputs? If consistent answers should emerge, it might be relatively straight-forward to infer about the impact of the program itself. 25/ A program involving a target population identified by socioeconomic characteristics, geographical location, or nutrition status can be regarded as 24/ See Annex, Section D. 25/ See Annex, Section C. - 31 - a special case. Such a program differs from those already discussed in that it is for a population identified by one or more of the nonprogram variables affecting the outcome. Hence, there is a built-in correlation between the measure of participation in the program and some other determinant(s) of the outcome. Most policy programs fall into this category. Impact measurement in such cases must be confined to the target population and involves essen- tially identical considerations to those relating to the nonexperimental programs discussed above. When within the target population, there is no (or not enough) variation in the program (input or utilization) variable, panel data. will be required to measure impact. For example, impact measure- ment in a particular intervention village of the Narangwal project is not possible if all that is known is that particular village was exposed to that intervention. However, when there is sufficient variance in the program variable, within the target population, this variance can be used to measure impact. Fcr example, the U.S. Food Stamp Program is designed for low-income families. Within that population, however, some families do not use the program at all and other families use the program to different degrees. This variation may permit estimating program impact. Within the perspective taken in the preceding paragraphs, an experimental approach is warranted when it is less costly, or when there is reason to believe that, given the available statistical techniques, panel data are insufficient to measure impact. Two basic situations may undermine the use of panel data. The first is when the underlying structural relation- ships are believed to change over time. Such a change may prohibit esti- mating the added effect of the program. The second situation is one where there is reason be believe that one or more of the nonprogram determinants - 32 - of the outcome will change during the monitoring period in a way that will confound any inference. Hence, experimental design may be used as a pre- cautionary measure against foreseen and unforeseen circumstances. However, taking the view that social experiments are costly, financially and other- wise, we must look for ways (a) to reduce their costs and (b) to maximize their usefulness. To reduce costs, we have to protect the measurement procedure against uncertainty rather than against ignorance because ignorance can be unnecessarily more costly. Uncertainty is defined here as a situation when one has good reason to believe that some known non-program determinants of the outcome may change significantly during the program and potentially have a confounding effect. Ignorance is defined here as a situation when one does not know the determinants of the outcome variable. In this case, one may wish to have an experiment by a random assignment of the intervention to "protect" the evaluation from all possible confounding effects. Unfortunately, this may be nearly impossible and unnecessarily costly. When considering the use of experimental design, the economic theory of the household and econometrics may help to reduce the "costs of ignorance" by suggesting which nonprogram variables might confound the evaluation and need to be controlled for. Alternatively, prior considerations should help to stratify and reduce the evaluation sample. 26/ The first step is to under- stand, possibly by some modelling, the behavioral and other determinants of a particular outcome. This effort, should be coupled with anthropological observations as well as a pilot survey. Once key non-program determinants 26/ This approach comes close to a "quasi experimental" design. - 33 - of the outcome have been identified, the number of variables for which the experiment has to be controlled may become manageable. The second step is to protect against uncertainty by limiting the control variables to those that may change during the life of the program. This step may further reduce the cost of the evaluation. Obviously, in reality the researcher confronts both uncertainty as well as ignorance. However, given the costs of evaluation research based on experimental design, the benefits from careful a priori modelling may be substantial. When experimental design is selected, an econometric approach to the data can help maximize its usefulness. As already mentioned, from a statistical viewpoint, impact measurement in a well-designed experiment can be satisfied just by comparing the means of the outcome variables between the treatment group(s) and the control group(s). This approach, however, reduces the acceptability and the universality of the results and has serious shortcomings in experimental contexts where policies are tested for future and wider applications. To increase the generality, or "external validity", of an evaluation based on experimental design, the researcher should include other variables or controls that affect the outcome measure and apply econometric techniques to the data. 27/ Thereby, one can test some basic behavioral hypotheses, and perhaps show explicity how a program relates to its environment. VI. CONCLUJSION The basic postulate of this paper is that although health related interventions aim to increase household and individual welfare, they may 27/ See Annex, Section E. - 34 - be inconsistent with the household's and individual's own objectives and opportunities. Where households can exercise choices vis-a-vis the program, studying these objectives and opportunities and relating them to particular interventions must be the first steps in understanding and predicting a program's potential impact. While analyzing household behavior is no mean task, it is nevertheless essential. By relating the concept of expected earnings to age-specific productivity, morbidity, and mortality proba- bilities, the economic theory of the household offers a conceptual framework in which to integrate relevant aspects of household behavior with the basic outcomes of health programs. Thereby, this theory can suggest testable hypotheses concerning program impact, and can help to identify various outcome measures for programs. Econometrics stresses the empirical aspects of the causal be- havioral relationships outlined by the economic theory and, given the social and financial costs of social experiments, can provide a useful and efficient framework for measuring program impact. An econometric approach can, at times, substitute for experimental design, and at other times, complement it, depending on financial and other costs as well as statistical con- siderations. When experimental design is warranted on statistical grounds because statistical control techniques are insufficient to measure impact, the economic theory of the household may helpt to minimize the cost of the experiment by suggesting which variables should be controlled for. Fur- thermore, once data from experimental design are available, the use of an econometric approach to explore those data can increase the credibility and generality of the results. - 35 - Thus, the combination of theoretical and other a priori considera- tions with econometric estimation techniques and experiments can prove the most effic:Lent way to measure program impact. - 36 - ANNEX A. A simple economic relationship--say between a household's caloric consumption (Y) and a vector X of household characteristics including, for example, household income--can be stated by Y = f (X). (1) This is an "exact" relationship in that it does not account for random ele- ments in human behavior. It is also a partial relationship in the sense that it may "ignore" other lines of causality; a more comprehensive underlying model may also deal, for example, with the effects of caloric consumption on household income. Given a relationship like (1) economic theory is usually concerned with the nature of the changes in Y due to changes in X, within a relevant range. B. Assuming that X represents just income and that Y and X are obser- vable, the econometric equivalent of relationship (1) can, for example, be loge (Yi) = a + b log e(X i) + vi, (i = 1 ---- n), (2) where the index i denotes one of n observations, and vi is a disturbance, or error, term which is usually assumed to be random, to have a normal distribu- tion with an expected value of zero and a constant variance, and to be uncor- related with other independent (right-hand) variables and with similar dis- turbance terms across observations. The term "a" is a shift parameter or "intercept", indicating levels of caloric consumption that are independent of X and v. The term "b" indicates, in this case, the change in log (Y) due to a unit change in log (X). Alternatively, "b" indicates the percentage change in Y due to a given percentage change in X, or the "elasticity" of Y with - 37 - respect to X. 28/ Assuming, no errors in the specification of (2), obtain- ing an unbiased estimate of "b" is possible when Xi is not correlated with vi 29/ C. The following underlying structure is generally assumed here for discussing measurement of the impact of a program. First, in a given en- vironment where such a program has been implemented, an outcome (Y) is a function of the household's utilization of the program (U) and par- ticular subset of household characteristics (X1), Y = f (U, X1) , (3) Where U and X may be dependent; that is, for biological and behavioral rea- sons, particular levels of program utilization may have different effects a2 in households of different characteristics; a ° aa1 Second, the utilization of the program is a function of the supply of program inputs (P) indicating the intensity by which the program reaches different households, and of a subset of household characteristics (X2) ; U = g (P, X2) ' (4) where XI and X2 are not necessarily mutually exclusive subsets at least conceptually. This structure can be reduced, by substituting (4) into (3), to Y = f[g(P,X2), X1] = h (P, X3), (5) where X3 iEi a set of household characteristics combining X1 and X2. 28/ Some other properties of this double-logarithmic function are discussed in section IV in the text. 29/ When Xi denotes a vector of variables, any two of those variables should also be uncorrelated. - 38 - Estimating relationship (5) may be sufficient to measure program impact. This alone is insufficient, however, for understanding this irnaet and thus for evaluating the program. In particular, such a relationship does not reveal whether variations in impact of given program inputs are due to different levels of utilization, to differences in impact of given :-lyitts on households of different characteristics, or both. The problem can be stated as the researcher's inability to test various hypotheses concerning X1 and X2 This problem can be solved by estimating relationship (4), which may be crucial for designing a cost effective program because this relationship identifies the users and levels of utilization of a particular program. 30/ D. A simple approach to panel data pertaining to a program is to pool baseline and program data, and estimate an explicit function of it g (t, Xit, Pi vit), (i = 1....n, t = 0,1) (6) where: t represents the time period a particular observation was obtained; 31/ Yi and Xi represent baseline or preintervention values; Yii = Yi + dYi and Xii = Xi + dXi represent post-intervention values; and Pi indicates program inputs. 32/ Assuming, without much loss of generality, that the relationship between caloric intakes and the other variables is linear within the relevant 30/ Although conceptually different, P and U are often interchanged and relationship (3) is estimated instead of (5). In such a case the researcher should be aware that he may attribute undue failure or success to program services. That is, the program may succeed or fail because of particular household characteristics; some program may succeed in one setting and fail in another. 31/ t = 0 for preintervention and t = I for during or post intervention. This illustration can be generalized to include more than two time periods. 32/ Alternative, Pi = 1 for participation in, or prevalence of, the program and P. = 0 otherwise. I - 39 - range, and that changes in income bring about identical changes in consump- tion over time, one can estimate this equation as illustrated in Figure 2. yit = a + bXit + b2t + b3pi + b4PiXit vit First, no interaction is assumed between the program (inputs) and household characteristics, b4 = 0; that is, program impact is independent of the level of household income. Curve (i) in Figure 2 indicates the general relationship between income and caloric consumption, or the pre-program relationship. Curve (ii) indicates the added effect of time, measured by b2 that is presumed to be independent of the other effects. For example, it can indicate that the effect of child growth on caloric consumption. Curve (iii) depicts the added effect of the intervention measured by b3' Second, we assume that Y measures child weight, for given age and sex, and that there is interaction between X and P, or b4 ' 0; that is, for a given level of program utilization the impact of the program, measured by gain in child weight, differs across households of different levels of income. For example, children of better-to-do families may be less exposed to diarrhea than children of poor families. Consequently, the children of relatively richer families may gain more in weight from a given school feeding program, ceteris paribus. This particular effect is depicted by line (IV) of the figure. Estimating program impact by using panel data is not feasible when "p" and "t" are statistically indistinguishable; that is, when some other events poterLtially affecting the outcome concur with the program. - 40 - y - (iv) Y-a+b2t+b3P+(bl+b4)X (- (iii) Y-a+b2t+b3P+blX -~ (ii) Y=a+b2t+blX (i) Y=a+b,X Figure 2 Figure 2 - 41 - E. Suppose that the sample underlying relationship (7) and Figure 2, is drawn from an experimental design which consists of n observations, m in the control group and (n-m) in the experimental or program group. That is, Pi = 0 for i = 1....m, and Pi = 1 for i = m+l1....n. To measure and test for intervention impact, the experimental design advocates measuring the difference E mdY dY - dY and testing for its statistical significance, where dY - 1 i e c c m is the mean change in the outcome in the control group, and dY is the corresponding statistic for the experimental group. This difference will reflect corretctly and fully program impact, only if the experimental design achieves complete randomization (or perfect matching) of all other potential effects; that is Pi does not correlate with any other determinant of the observed outcome. To achieve the same goal an econometrician would estimate relation- ship (7), by either .. Y it G + b3 Pi + Vit (8) It or dY. E + b3 P + Wi where wi is El disturbance term similar to vi- It can be shown that I,. dY - dY = b b b e c 3 3 3 (10) when Pi does not correlate with any of the right-hand variables in equation (7). Hence, while econometric and experimental approaches should yield identical statistical results because of the underlying statistical rationale, the econometric approach specifies particular structural and functional re- lationships and tries to account for all relevant determinants of the outcome. - 42 - BIBLIOGRAPHY 1. Basta, Samir S. and Churchill, Anthony. Iron Deficiency Anemia and the Productivity of Adult Males in Indonesia. Washington, D.C., World Bank, Staff Working Paper No. 175, 1974. 2. Becker, Gary S. "The Theory of Allocation of Time". Economic Journal, 75: 493-517 (1965). 3. Bennet, Carl A., and Lumsdaine, Arthur A. Evaluation and Experiment, Some Critical Issues in Assessing Social Programs. New York, Academic Press, Inc., 1975. 4. Campbell, Donald T. and Stanley, Julian C. Experimental and Quasi- Experimental Designs for Research, Chicago, Rand McNally College Pub- lishing Company, 1966. 5. Chernichovsky, Dov. "The Demand for Nutrition." Washington, D.C., World Bank, 1977. (Mimeographed.) 6. Chernichovsky, Dov, and Coate, Douglas. The Choice of Diet for Young Children and Its Relation to Children's Growth. New York, National Bureau of Economic Research, Working Paper No. 219, 1977. 7. Chernichovsky, Dov, and Kielman, A. Aanfried. "Socio-Economic Status, Diet and Preschool Child Growth in Rural Punjab-India." Baltimore, The Johns Hopkins University, 1977. (Mimeographed.) 8. Christakis, George, ed. Nutritional Assessment in Health Programs. Washington, D.C., American Public Health Association, 1974. 9. Cuca, Roberto, and Pierce, Catherine. Experiments in Family Planning. Baltimore-and London, The Johns Hopkins University Press for the World Bank, 1977. 10. Eckholm, Erik, and Record, Frank. The Two Faces of Malnutrition. Washington, D.C., World Watch Institute, Paper No. 9. 11. Duncan, Otis Dudley. Introduction to Structural Equation Models. New York, Academic Press, 1975. 12. Fuchs, Victor R. Who Shall Live? Health, Economics and Social Change, New York, Basic Books Inc., 1974. 13. Grossman, Michael. The Demand for Health: A Theoretical and Empirical Investigation. New York and London, Columbia University Press for National Bureau of Economic Research, 1972. - 43 - 14. Grossman, Michael, and Edwards, Linda Nasif. An Economic Analysis of Children's Health and Intellectual Development. New York, National Bureau of Economic Research, Working Paper No. 180, 1977. 15. Johnson,, J. Econometric Methods. 2nd ed., New York, McGraw-Hill Book Company, 1972. 16. Lancaster, Kelvin J. "A New Approach to Consumer Theory". Journal of Political Economy, 74: 132-57 (1966). 17. Latham, L., Latham, M., and Basta, Samir S. The Nutritional and Economic Implicat:ions of Ascaric Infection in Kenya. Washington, D.C., World Bank, Staff Working Paper No. 271, 1977. 18. Lechtig, Aaron. "Effect of Improved Nutrition During Pregnancy and Lactation on Developmental Retardation and Infant Mortality". In: P.L. White and N. Selvey, ed., Proceedings of Western Hemisphere Nutrition Congress, Massachusetts, Publishing Sciences Group, Inc., 1975, p. 117. 19. Mincer, Jacob. Schooling, Experience and Earnings. New York and London, Columbia University Press for National Bureau of Economic Research, 1974. 20. Mushkin, Selma. "Health as an Investment." Journal of Political Economy, 70: 129S-157 (1962). 21. Nasoeticn, Andi Hakim. "Spurious Correlation as a Result of Constraints in Randomization". Research Methodology, Agricultural Development Teaching Forum Council, 46: (1975). 22. Reutlinger, Shlomo, and Selowsky, Marcelo. Malnutrition and Poverty. Baltimore and London, The Johns Hopkins University Press for the World Bank, 1976. 23. Ross, John and Smith, Perry. Orthodox Experimental Designs. In: H.M. Blalock and A.B. Blalock, ed., Methodology in Social Research, New York, McGraw-Hill, 1968, p. 339. 24. Selowsky, Marcelo. "A Note on Preschool-age Investment in Human Capital in Developing Countries". Economic Development and Cultural Change, 24: 707-20 (1976). 25. Selowsky, Marcelo, and Taylor, Lance. "The Economics of Malnourished Children: An Example of Disinvestment in Human Capital". Economic Development and Cultural Change, 22: 17-36 (1973). 26. Smucker, Celest Mary. Socio-Economic and Demographic Correlates of Infant and Child Mortality in India. Ann Arbor, Michigan, University of Michigan, Phd. Dissertation, 1976. - 44 - 27. Snedecor, George W., and Cochran, William G. Statistical Methods, 6th ed., Ames, Iowa State University Press. 28. Stoch, M. B., and Smythe, P. M. "15-year Development Study on Effects of Severe Undernutrition During Infancy on Subsequent Physical Growth and Intellectual Functioning". Archives of Disease in Childhood, 51: 326-36 (1976). 29. Taylor, E. Carl, et al "Malnutrition, Infection, Growth and Development: The Narangwal Experience". Baltimore, The World Bank and The Johns Hopkins University, 1978. (mimeographed.) 30. U.S. Department of Health, Education, and Welfare. Summary Report: New Jersey Graduated Work Incentive Experiment. Princeton, New Jersey, 1973. 31. World Bank. Health Sector Policy Paper. Washington, D. C., World Bank, 1975. RECENT PAPERS IN THIS SERIES: No. TITLE OF PAPER AUTHOR 275 Land Reform in Latin America: Bolivia, Chile S. Eckstein, G. Donald Mexico, Peru and Venezuela D. Horton, T. Carrol (consultants) 276 A Model of Agricultural Production and Trade R. Norton. in Central America C. Cappi, L. Fletcher, C. Pomareda, M. Wainer (consultants) 277 The Impact of Agricultural Price Policies on M. Osterrieth, Demand and Supply, Incomes and Imports: E. Verrydt, J. Waelbroeck An Experimental Model for South Asia (consultants) 278 Labor MaLrket Segmentation and the Determination D. Mazumdar of Earnings: A Case Study M. Ahmed 279 India - Occasional Papers M. Ahluwalia, J. Wall S. Reutlinger, M. Wolf R. Cassen (consultant) 280 Educational Effects of Class Size W.D. Haddad 281 Relieving Traffic Congestion: The Singapore P.L. Watson Area License Scheme E.P. Holland 282 World Trade and the International Economy: B. Balassa Trends, Prospects and Policies 283 Urban Land Policy Issues and Opportunities H. Dunkerley, A. Walters, J. Courtney. W. Doebele, D. Shoup, M.D. Rivkin (consultants) 284 Pakistan: Forestry Sector Survey S.A. Draper, A.J. Ewing. J. Burley, G. Grayum (consultants) 285 The Leisure Cost of Electric Power Failures M. Munasinghe 286 Shadow Pricing and Power Tariff Policy M. Munasinghe, J. Warford 287 Wages Capital Rental Values and Relative S. Guisinger (consultant) Factor ]?rices in Pakistan 288 Educational Reform in the Soviet Union: I. Blumenthal, Implications for Developing Countries C. Benson (consultants) No. TITLE OF PAPER AUTHOR 289 Petroleum and Gas in Non-OPEC Developing R. Vedavalli Countries: 1976-1985 290 Major Reforms of the Swedish Education System A. Heidenheimer (consultant) 291 Industrialization, Technology and Employment - T.G. Rawski (consultant) China 292 Development and Income Distribution - Zambia C. Blitzer 293 World Potash Survey W. Sheldrick, H. Stier 294 The Economic Dimensions of Malnutrition in M. Selowsky Young Children 295 The Technology of Rural Development J.P. McInerney (consultant) 296 The Financial Cost of Agricultural Credit: C.D. Datey (consultant) A Case Study of Indian Experience 297 Agricultural Sector Planning Models: A.C. Egbert A Selected Summary and Critique 298 Textbooks and Achievement: What We Know S.P. Heyneman. J.P. Farrell, A. Sepulveda-Stuardo (consultants) 299 An Economic and Social Analysis of the C. Bruce Chao Phya Irrigation Improvement Project II Y. Kimaro 300 Two Studies of Development in Sub-Saharan S. Acharya Africa B. Johnston (consultant) 301 The Intermediate Sector, Unemployment, and W.F. Steel The Employment-Output Conflict: A Multi-Sector Y. Takagi (consultant) Model PUB HG3881.5 .W57 W67 no.302 Chernichovsky, Dov. The economic theory of the -ho.eho1d and impact PUB HG3881.5 .W57 W67 no.302 Chernichovsky, Dov. Theweeconomic theory of the household and impact 'measurement of nutrition and