Tim Miller > Curriculum Vitae > Malnutrition and mortality among Bolivian children > Chapter Five

Child mortality rates in Bolivia have been falling quite rapidly over the last 15 years. In Chapter 3, macro evidence on urbanization, increases in GDP per capita, improvements in nutrition, growth of primary health care, fertility changes, and expansion of the educational system were considered as possible explanations of this rapid decline. Of these factors, only the expansion of the educational system seems plausible. A causal interpretation of this macro-level relationship must rest upon an understanding of the presumed micro-level relationship between education and infant health. In this chapter I use a proximate determinants framework to examine the impact of education on early child mortality in Bolivia at the micro-level using information on children collected from about six thousand Bolivian women in June, July, and August of 1989. Examining this relationship at the micro-level can provide answers to two important questions. First, to what extent is the association between maternal education and child mortality robust to consideration of other socio-economic factors? Second, assuming that there is some association net of these factors, what are the behavioral mechanisms responsible for this?

Large differences in child mortality by maternal education have been observed in most countries throughout the world. Evidence collected in 41 countries as part of the World Fertility Survey shows that children of mothers with no formal education face substantially higher risks of infant death. These data are presented graphically in Figure 5.1. These differentials are quite large -- with children of mothers with no formal education facing up to twice the risk of children of more educated mothers (7+ years). Evidence presented in a comparative analysis of 11 countries using DHS data (Bicego and Boerma (1991)) also show a large educational differential in infant mortality. Additionally, as seen in Figure 5.2, this differential is generally greater in the postneonatal than neonatal period. Since postneonatal mortality is typically much greater than neonatal mortality, these relative differentials represent very large absolute differences in mortality rates.

In most societies, being more educated also means being wealthier. In fact, in the absence of income measures, educational attainment is often used as a proxy for wealth. To what extent is the universally observed relationship between maternal education and child survival partly or entirely explained by the effect of wealth?

Because of the paucity of income data in developing countries, it is extremely difficult to separate wealth from education effects. Generally, several proxies are used for wealth: father's educational status, father's occupation, housing characteristics (e.g., floor or roof materials, presence of plumbing), and household possessions (e.g., tractors, radios, televisions, bicycles). For example, in a comparative analysis of DHS data, Bicego and Boerma control for income using an index of household economic status: a six-point scale based on two points for TV or one point for radio plus two points for non-dirt floor plus two points for car, tractor, or truck or one point for motorcycle. Given the data constraints, these studies provide a rough estimate of presumed wealth effects. Cleland

and van Ginnekan (1990) summarize the results of several such studies, concluding that on average about half of the observed effect between maternal education and child survival is an income effect. In Bicego and Boerma's 11-country study, the income effects accounted for between 30 and 40 percent of the observed relationship between maternal education and child survival. This evidence indicates that there remains a strong association between maternal education and child survival independent of these crude measures of wealth.

The question as to what underlies this association has been raised by several authors beginning with Caldwell (1979), who observed in his Nigerian study that maternal education -- much more so than paternal education or occupation -- was strongly associated with increased child survival. He gave three possible reasons for this. First, mothers become less fatalistic, break with tradition and adopt new child care behaviors. Second, educated mothers become more capable of manipulating the modern world. Third and most importantly, there is a change in the traditional balance of family relations in a family in which the mother is educated. This change leads to "child-centered" families.

Schultz (1984) posits five possible explanations in presenting his general framework for the analysis of mortality. First, education may increase the productivity of health inputs. For example, educated mothers may know to boil water in order to kill water-borne pathogens. Second, it may reduce costs of information about the optimal use of health inputs. Where information about what is "best" is scarce, educated mothers may be at an advantage in seeking out such information. Third, education may increase family income. Fourth, education may increase the mother's time costs. To the extent that mother's time is an input in child health, such a result would serve to decrease child health. Fifth, education may change preferences for child health and family size.

In this chapter, data limitations force us to consider only two types of mechanisms: the adoption of more healthful behaviors and the greater efficacy of these behaviors among more highly educated women. The proximate determinants framework provides a means for assessing these mechanisms. This approach was first devised by Bongaarts to trace causal relationships in fertility. Its adaptation for the study of mortality was first proposed by Mosley and Chen (1984). The essential feature of such models is the formulation of a set of intermediate bio-medical variables (proximate determinants) through which the ecological, cultural, and socio-economic forces affect mortality. (See Figure 5.3). For example, better educated mothers may have lower infant mortality because of healthier reproductive patterns or greater use of prenatal care. The promise of the proximate determinant model lies in its ability to give bio-medical explanations for correlations observed between socio-economic factors and child mortality.

The proximate determinants framework has been integrated into an economic choice model by Schultz (1985). There are two sets of equations in this proximate determinants model. The first set predicts the various proximate determinants as a function of social, cultural, and economic factors. This set is known as the behavioral equations or the demand equations. There is one equation for each of the proximate determinants. The second equation predicts the health outcome (usually mortality) as a function of the

proximate determinants. This is known as the biological equation or the health production function equation.

A major research problem in health analysis is unobserved health heterogeneity. To the extent that subjects alter health behaviors in response to some personal health factors which the researcher cannot observe, estimates of the relationship between health behaviors and outcomes will be biased. For example, if women in poor health seek out prenatal care, research is likely to reveal a positive relationship between early prenatal care and poor health outcomes. There are two methods for dealing with unobserved health heterogeneity. The first relies on a two-stage estimation technique (Schultz, 1984). All variables are constructed so that they are exogenous or predetermined with respect to their influence on health outcomes. Endogenous variables -- e.g., the seeking of prenatal care -- are replaced by predicted values based on exogenous factors thought to influence health behaviors. In this way, the predicted endogenous variable reflects the hypothetical behavior of an individual who was unaware of his/her personal health heterogeneity.

An alternative approach is to rely on motivational data to reveal unobserved heterogeneity. In the prenatal care example cited above, motivational data as to why women sought out prenatal care would allow the researcher to control for unobserved health heterogeneity. Such motivational data are becoming increasingly available in health surveys.

The two-stage estimation approach is unlikely to have useful application in health policy research -- particularly research focused on developing countries. First, there are not many data sets in which contain enough instruments to make first-stage estimation possible. Second, understanding motivation eliminates the need for this technique. It is easier to collect information on why a person went to prenatal care than to collect the necessary price and income data needed for estimation of this behavior. Third, in addition to being easier to collect, motivational data is generally more convincing than two-stage estimation results. Indeed, in studies utilizing two-stage techniques, the change in parameter estimates brought about by this technique are usually justified by appealing to beliefs about motivation.

Path analysis differs from two-stage estimation in that the biological equation is estimated using the true values of the proximate determinants rather than the those values predicted using a set of background variables. Thus, path analysis implicitly assumes that there is not a significant amount of unobserved health heterogeneity. The impact of the cultural, economic, and social factors can be traced via the pathways observed through the behavioral and biological equations. For example, one could examine the impact of education on mortality though its effect on breast-feeding behavior and on health service utilization. The impact of a particular channel is simply the product of the estimated behavioral coefficient linking education with increased use of health services and the estimated biological coefficient measuring the effect of health service usage on mortality. The total impact is the sum of all pathways.

Both path analysis and two-stage estimation share a common weakness which has prevented either from becoming widely used. Both rely on the existence of a complete set of proximate determinants through which -- and only through which -- the social, economic, and cultural factors influence mortality. Not much information about medical and non-medical health behaviors has been collected in large-scale health interview surveys. The information usually collected pertains to medical processes before and at birth rather than during the post-neonatal period where most child deaths occur. For example, in the DHS, information is collected on the mother's tetanus immunization, her prenatal visits, and assistance at delivery.

When faced with an incomplete list of proximate determinants, most researchers have chosen to estimate a hybrid model in which bio-medical and socio-economic characteristics are mixed together. (See for example, Barbieri (1990) and Bicego and Boerma (1991)). Such models have been criticized by the proponents of two-stage estimation primarily on two grounds. First, the parameter estimates of the biomedical factors will be biased. Second, the parameter estimates of the socio-economic factors are easily misinterpreted. (Da Vanzo and Gertler, 1991).

I do not share this skepticism. The first is not a valid reason for preferring path models over hybrid models. Both would be subject to biases caused by unobserved heterogeneity. The bias does not result from the mixture of biomedical and socio-economic variables in a hybrid model. It is, however, a powerful argument in favor of two-stage estimation which as discussed earlier can overcome the biases associated with unobserved health heterogeneity. In those rare circumstances in which the data permit such estimation, it is the preferred methodology. The second criticism is that the parameter estimates for the socio-economic factors in a hybrid model may be misinterpreted as representing the full impact of those factors on mortality. Hybrid models allow one to examine the indirect effects of the socioeconomic variables on mortality. The full impact also includes an indirect effect via the bio-medical factors also included in the hybrid model. This is an issue of interpretation and not a flaw of hybrid models.

Hybrid models also allow for the examination of interaction effects between socio-economic factors and behaviors. In addition to using different amounts of health inputs, more educated women may use health inputs differently. This means that there are likely to be important interaction effects between education and public program interventions. Some of these program efforts will be of greater benefit to more highly educated women, others will be of greater benefit to women with little of no schooling. The evidence is mixed as to whether program interventions serve to widen or narrow educational differences in child mortality. This effect may depend on the characteristics of the health intervention.

Caldwell (1979) found that the presence of a hospital served to increase educational difference in mortality. He attributed this to the greater ability of higher educated women to manipulate a bureaucratic environment to their and their children's advantage. By contrast, Rozensweig and Schultz (1982) in their study of child mortality in Colombia find that access to public and private medical facilities in urban areas serves to reduce educational differentials in mortality. They view the public health information component of interventions to be a substitute for the specific knowledge and skills of higher educated mothers.

Public health programs aimed at increasing access to water can also serve to widen or to narrow educational differences in child mortality as noted by Esrey and Habicht (1988) in their analysis of the Malaysian Family Life Survey. An increase in water quantity should benefit literate mothers more since they are more likely to be practicing better hygiene (e.g., hand washing after defecating). However, an increase in water quality should be of more benefit to illiterate women since literate women presumably know to boil water to improve water quality. They find that the beneficial effect of piped water is larger among literate women and conclude that health improvements linked to provision of water came mainly from increased quantity of water.

This analysis will try to answer two important questions about maternal education and child survival. First, does maternal education have an effect on child mortality independent of husband's education, occupation, and other socio-economic factors? Second, what are the bio-medical or behavioral factors underlying the association between maternal education and child survival?

This analysis is divided into two parts. In the first part, neonatal and postneonatal mortality are analyzed. Four types of equations will be estimated (see Figure 5.4). The first set are reduced-form equations which measure the impact of the socioeconomic, cultural, and ecological factors on mortality. Here the effects of maternal education net of other socio-economic factors are examined. The second is a bio-medical equation (a health production function) which examines the impact of the proximate determinants on mortality. The third equation is a hybrid equation which estimates the impact of the proximate determinants and maternal education. The fourth set is a path analysis which utilizes the hybrid equation to assess the direct and indirect impact of maternal education on child mortality. A two-stage estimation of the health production function was unsuccessful due to the lack of suitable exogenous variables required for first-stage estimation.

The second part is an analysis of cause of death in the neonatal and postneonatal periods in which the hybrid equation used in part one is re-estimated. Examination of child mortality by cause will provide a better basis for causal explanations of the direct and indirect impact of maternal education on child mortality.

The risks an infant faces during birth and the first month of life are very different from those faced after this period. Health researchers typically separate the infant mortality rate into two parts: the neonatal period representing the first 28 days of life and the post-neonatal period representing the remainder of the first year of life. With developing country data, this division is difficult. First, we often lack detailed information on the timing of infant death that would enable such a separation. Second, even when such information is available it is often questionable due to misreporting.

In the DHS for Bolivia we have information on the age at death of children reported in days, months, or years. This reporting is subject to age-heaping problems at ages 6,12,18, and 24 months as seen in Figure 5.5. The most serious heaping problem occurs at 12 months. This presents a problem for the demographer interested in reporting an infant mortality rate. We believe that some portion of these deaths occurred before the 12th month of life and hence should be included in a calculation of the risk of infant mortality. When did these misreported deaths actually occur?

For this analysis of the risk factors for neonatal and post-neonatal mortality, I follow the example of researchers Bicego and Boerma (1992), who chose to examine the neonatal period and a post-neonatal period defined as a period up to age 2, rather than age 1. In this way, the question of heaping of deaths is finessed. Though we are not sure whether those deaths heaped at age 12 months occurred before age 1 or after, we are reasonably sure they occurred before age 2. Thus, the risk of death in the first month of life (1q0) and risk of death in the next 23 months of life (23q1) will be analyzed.

A total of 7,923 women between the ages of 15 and 49 years were interviewed in June, July, and August of 1989 as part of the Bolivian Demographic and Health Survey. Roughly 70% of these women reported ever having given birth. These 5,542 women gave basic information on a total of 22,338 births. My analysis is limited in this section to births in the last five years (5,784 births) for two main reasons. First, we have more detailed health information on these children (e.g., reported size at birth, place of delivery, maternal tetanus immunization, prenatal care received by mother). Second, much of the information we have is current status information that we are only certain pertains to the most recently born children. For example, information about where the family is currently living may not pertain to older children because the family may have moved since that older child's birth.

Of the 5,784 children born in the last five years, about 91% were alive at the time of the survey. For the 9% who died (537 children) the mothers were asked for detailed information about the age at death and cause of death. I examine two samples in this analysis. The first is used for estimating the risk of death before age 2 (23q0) and the risk of post-neonatal death (23q1). It consists of the 3,292 children who were born at least 24 months before the survey date. In this sample, 10.5% of the children (346 children) die before age 2. Of these deaths, 110 occur in the first month of life and 236 occur in the next 23 months. This defines a probability of death in the neonatal period of 33.4 per thousand (110 deaths/3292 births) and a probability of death in the post-neonatal period of 74.2 per thousand (236/3182 one-month-olds).

The second sample is used for estimating the risk of death in the neonate period (1q0). It consists of the 5,635 children born at least a month before the survey. In this sample, 179 children die in the first month of life. Thus, the probability of death in the neonate period for this sample is 31.8 per thousand (179/5635).

There are two types of variables used in this analysis: background factors and proximate determinants. There are three sets of background variables: maternal education, paternal characteristics, and family and residential factors. (See Table 5.1).

There are 5 sets of proximate determinants: reproduction factors, child characteristics, prenatal and natal medical care, community medical care, and household water and sanitation. (See Table 5.2)

The first set of background factors analyzed is the educational attainment of the mother. This is coded as a series of four dummy variables: educ0, educ1, educ2, and educ3 representing mothers with no schooling (19% of all births), those who have some

education (basico: 1-5 years -- 46%), those with at least intermedio level schooling (6-8 years -- 14%), and those with 9 or more years of schooling (21%).

The second set of background variables are father's education and occupation. These are introduced as proxies for household income -- which is unmeasured in the DHS survey. Father's education is coded as a series of four dummy variables: educ0, educ1, educ2 , and educ3 representing fathers with no schooling (6% of all births), those who have some education (basico: 1-5 years -- 36%), and those with at least intermedio level schooling (6-8 years --17%), and those with 9 or more years of schooling (37%). The variable heduc99 indicates that there is no father present in the household (3% of the births). Father's occupation is coded as a series of four dummy variables: oc1, oc2, oc3, and oc4 representing a four class division of occupations: professional/sales/clerical (20%); agricultural (36%); domestic/informal sector services (12%); and manual labor (28%);.

The third set of background variables represent family and residential characteristics. The variable indig which indicates whether the family usually speaks an indigenous language at home (21% of the births). Residential characteristics are represented in the variables rural, altip, valley, and tropics -- which reflect if the respondent lives in a rural area (48%) and in which of the three major ecological regions they live. The variable altitude represents the elevation of the city, town, or village in which the respondent lives.

There are five sets of proximate determinants in this model: reproductive factors, child characteristics, prenatal and natal medical care, community medical care, and household water and sanitation.

In reproductive factors, I have controlled for parity, birth-spacing, maternal age, and year of child birth. I have coded parity in two dummy variables: first representing firstborn children (19% of the births) and six representing children sixth-born or greater (24%). Birth-spacing is measured by the variable short representing children born after a birth interval of less than 24 months (28%). Maternal age is coded in a series of five-year wide age intervals: m15 to m45. In addition, I have proxied for the year of birth (b84, b85, b86). As mortality is rapidly declining in Bolivia, there may be significant differences in mortality over time.

The second set of proximate determinants are child characteristics such as boy, twin, and birth weight. Boys (coded as male) are generally believed to be at higher risk than girls due to congenital weakness of male babies. Twins (coded as twin -- 1% of the births) are generally believed to be high risk deliveries. Most importantly, low birth weight babies have been shown to be at much higher risk of mortality. Low birth weight indicates premature delivery, intrauterine growth retardation, or both. These factors place children at significantly higher mortality risk. We lack data on birth weights in the DHS. Instead, mothers were asked whether in their opinion the child was very small at birth, smaller than normal, normal, larger than normal, or very large. Of all these responses, I found only the first ("very small" coded as vsmall -- 9% of all births) to have a significant effect on mortality risk relative to all other reported birth sizes (smaller than normal, normal, larger than normal, much larger than normal).

There are three variables which measure prenatal and natal medical care. Two variables measure prenatal medical care: prenate which indicates the mother received prenatal care (49% of the births) and tetanus which indicates the mother received at least one tetanus immunization (23%). The variable delivery indicates the mother delivered this child in a medical facility rather than at home (40%). Unfortunately, I cannot include any information on post-natal medical care in these models. This is because such information was only collected for living children in the DHS-I Surveys. One post-natal medical care of particular interest is immunization. We suspect that children who were not immunized are at much higher risk of mortality; unfortunately we do not know the immunization status of those children who have died. One way around this is to impute this information. Such attempts are unlikely to be successful mainly because mothers behave differently toward their children: immunizing some and not others. Furthermore, child death may increase the likelihood of subsequent immunization of siblings -- resulting in a positive association between immunization imputed from siblings and child death. (Diamond, et.al., 1991). In DHS-II Surveys, immunization information was collected on all children -- living or dead.

The fourth set of proximate determinants are measures of community access to medical care. Two variables indicate those communities with immunization coverage below 50% for DPT (coded as lodpt -- 8% of births) and for measles (coded as lomeas -- 17%). The other indicates communities in which no child has been delivered in a medical facility and no child has ever been taken to a medical facility for treatment of diarrhea (usecat1). I think these communities are most likely in remote locations with no nearby health facilities (6% of all births).

The fifth set of proximate determinants measure household water and sanitation. I measure water supply by the variable piped which indicates the household has piped water (71% of all births). I introduce two sanitation variables: latrine indicating the household has a latrine (23%) and toilet indicating the household has either a septic tank or a sewer line hookup (23%). The excluded category is the lack of any sanitary facilities (54%).

Breast feeding is not examined in this model. Breast feeding is thought to be an important determinants of child health. Breast milk is a source of complete nutrition for the first six months of life and it provides immune protein substances which help protect the infant against infections which enter the body through the intestinal tract (Palloni and Tienda, 1986). Therefore, it is expected that failure to begin breast feeding or premature weaning may lead to increased risk of child death. It is difficult to assess the effects of breast feeding using the logistic model chosen in this analysis (discussed below). Reverse causality is a serious problem. That is, child death will lead to early cessation of breast feeding and hence a strong relationship will be observed between early weaning and child mortality. The problem of reverse causality can be mitigated by the use of a hazards model (Pebley and Stupp (1987)) or use of a logistic model of several periods by examining breast feeding status in the previous interval. (Palloni and Tienda (1986)). This multi-period logistic approach is not viable in the simple two period model which will be examined here.

In Bolivia, early weaning is probably not an important determinant of differences in child mortality. First, most Bolivian women breastfed for over a year and therefore, the majority (75%) of the children who died in the first two years of life had not yet been weaned. Second, those children (25%) who were weaned prior to death do not appear to have been weaned significantly earlier than other children (10 months versus 12 months, respectively). Therefore, choice of the logistic model and the exclusion of breast feeding as a proximate determinant should not bias the conclusions presented here.

Child death can be modeled using binomial logistic regression analysis. Binomial logistic models are part of the general class of discrete regression models that are used for events that assume discrete rather than continuous values. The binomial logit is used to model events that can take on only two values -- in this case: life or death. The logit takes its name from the logit transformation which is simply the log of the probability of the event divided by the probability of the non-event. Risk factors in mortality can be assumed to have a multiplicative effect on the probability of death rather than an additive effect -- which insures that the mortality risk is bounded by 0 and 1. When the mortality risk is a multiplicative sum, the logit-transformation of this risk is a dependent variable linear in parameters and easily estimated by maximum likelihood techniques. The predicted value of the untransformed dependent variable can be interpreted as the probability that the observation is an event, in this case that a particular child died. It ranges from 0 to 1. The exponentiated value for a parameter coefficient is referred to as the relative odds ratio -- measuring the increase in risk associated with the independent variable of interest. For example, an odds ratio of 2.0 means an individual with characteristic X faces approximately twice the risk faced by those without characteristic X, all other things being equal. Thus, the magnitude of these parameter estimates approximates the relative risk faced by a particular group. For a general discussion of the logit regression equation the reader is referred to Maddala (1983, Chapter 2) or Hosmer and Lemeshow (1989). For this analysis, I make use of the statistical languages mainframe SAS Version 5.0 in the CMS environment and mainframe TSP version 4.2 in the UNIX environment.

Three reduced-form equations are estimated. The coefficients for education in these three equations are graphed in Figure 5.7 for neonatal mortality and 5.8 for postneonatal mortality. The regression results are presented in Tables 5.3 and 5.4. In the first equation, we examine the gross effects of education on mortality. Here we observe that the children of less educated women face substantially higher mortality risks in both the neonatal and postneonatal periods. For example, children of women with no or limited (less than 5 years) schooling face twice the mortality risks of better educated (more than 8 years) women in both periods. While the relative risks associated with maternal education are roughly the same in both periods, the absolute effects are much greater in the post-neonatal periods -- since postneonatal mortality is two and a half times greater than neonatal mortality.

The second reduced-form equation examines the effects of maternal education after controlling for household income. As noted earlier, there are no income measures in the DHS -- nor in most health surveys conducted in developing countries. Instead, the occupation and education of the husband are frequently used as proxies. After controlling for husband's education and occupation, maternal education continues to have a strong effect on both neonatal and postneonatal mortality. In the neonatal period, there is hardly any change in effects -- the net effect of maternal education controlling for paternal education and occupation is nearly identical to the gross effect. In the postneonatal period, the net effect of maternal education is between 70% and 80% of the gross effect. As noted earlier, previous studies have found estimated the net effect to be between 50% and 70% after controlling for presumed wealth effects.

The third reduced-form equation includes language usually spoken by the family, urban location, region, and altitude information. A strong effect of maternal education on

neonatal and postneonatal mortality remains after controlling for these effects. It appears that the association between maternal education and child mortality is robust to consideration of other socio-economic factors. We now turn to examination of the possible behavioral mechanisms which underlie this socio-economic correlation.

I estimate two proximate determinants models: one in the neonatal period, the other in the postneonatal period. The results of these two estimations are presented in Table 5.5 and 5.6. The discussion will focus on the relative and absolute effects of the proximate determinants -- comparing the neonatal and postneonatal periods. Many of the parameters have large standard errors and extreme caution should be exercised in interpreting these parameters.

The first set of proximate determinants control for reproductive factors: parity, preceding birth interval, mother's age, and year of birth. While the relative effects of parity and birth spacing are similar in the both periods, the absolute effects are much greater in the post-neonatal period. High parity children (6 or more) and firstborn children both face elevated mortality risks. Children born after a short preceding birth interval are seen to face between 1.7 and 1.9 the risk of children born after a long birth interval. Not much effect is seen due to mother's age -- except in the postneonatal period in which children of mothers aged 35 to 44 face substantially lower risks than children of mothers aged 30-34. Mortality is falling rapidly in Bolivia and one might therefore expect to find a significant effect of timing of birth. No such effect is observed -- probably due to small sample size.

In examining child characteristics, we note that twins face high relative risks in both periods. Children who were judged to have been much smaller than normal have very high relative risks in the neonatal period (3.75). While the absolute risks in both periods are

quite similar, boys have higher relative risks in the neonatal period. This is attributable to the greater congenital weakness of male babies. A large absolute effect for boys is estimated in the postneonatal period -- but we have little confidence in this estimate due to the large standard error. That twins face similarly high risks of both neonatal and post-neonatal mortality is unexpected and indicates that twins face high mortality risks not merely because of problems at delivery. They are also at greater risk of dying in the post-neonatal period -- even after controlling for low birthweight. The sibling competition argument which attempts to explain higher mortality of closely spaced births is also a persuasive argument for twin mortality. In this case, the primary competition may be for mother's milk.

Examining medical care factors, we see that relative risks associated with these factors are similar in the two periods -- but that there are dramatically larger absolute effects in the postneonatal period. The strong role played by prenatal care and tetanus immunization on infant mortality is well-known. This has usually been interpreted as demonstrating the important role played by these medical interventions -- in the neonatal period. The fact that the majority of this beneficial effect occurs in the postneonatal period suggests that these factors may be proxies for some other behaviors or characteristics.

No information is collected as part of the DHS on the immunization status of children who have died. I have attempted to measure the effects of immunization through two variables: lodpt and lomeas. These measure those communities with less than 50% immunized children for DPT and measles. No information was collected in the Bolivian DHS as to the location of the nearest clinic. I attempted to control for access to health care by the variable -- noclinic. This represents those communities in which no child was taken to a medical facility for treatment of diarrhea, nor any children delivered in a medical facility. None of the parameter estimates for these three variables is statistically significantly different from zero. I think this is a reflection of the poorness of these proxies rather than demonstrating that immunizations and health facilities do not have strong effects on child mortality.

The impact of sanitation is measured in three variables: presence of latrine, toilet, and piped water. Neither estimates of the effects of piped water nor of latrine are statistically significant from zero. However, presence of a toilet is seen to have strong effects in both periods. Sanitation variables (such as latrine and toilet) are sometimes used as proximate determinants measuring direct health benefits from reduced exposure to pathogens, but are also frequently used as proxies for wealth. Both interpretations are plausible.

Problem in two-stage estimation

In order to control for possible unobserved heterogeneity, two-stage estimation was attempted. Three measures of prenatal and natal medical care were estimated using the background variables listed in Table 5.1 as instruments. The results are presented in Appendix 5.1. The background variables are not very good predictors of these behaviors. The parameter estimates derived in the second-stage are all insignificant -- with rather large standard errors. Further, there is a good deal of colinearity in the predicted probabilities of receiving prenatal care and of receiving delivery assistance. This failure in estimation can be attributed to the paucity of good instruments in the Bolivian DHS.

Do the proximate determinants mediate the effect of maternal education on mortality? To test this, a hybrid equation is estimated in which maternal education is included with the proximate determinants. These estimations are presented in Tables 5.7 and 5.8. In Figure 5.9, the estimated association between maternal education and neonatal mortality net of

other socio-economic factors (as derived in the reduced-form equations) is compared with the estimated effects derived in this hybrid equation. The figure shows that most of the educational differences in neonatal mortality are mediated by the proximate determinants.

In contrast, very little of the effect of maternal education on postneonatal mortality is mediated by the proximate determinants. (See Figure 5.10). This is not too surprising in that we are missing important proximate determinants for the postneonatal period, such as domestic child care practices.

As outlined earlier, several previous studies have found important interaction effects between maternal education and bio-medical factors. I attempted to estimate several such interaction effects in the hybrid model: education with prenatal care, tetanus immunization, assistance at delivery, lack of clinics, presence of a latrine, presence of a toilet, and birth-spacing. As seen in the Appendix 5.2, none of these effects is statistically significant, perhaps due to the small sample size.

Now, we turn to an examination of the direct and indirect effects of maternal education in the neonatal and postneonatal period. This consists of two sets of estimations. In the first-stage, the effects of all the background factors (maternal education, paternal characteristics, family language, and residential location factors) on each of the proximate determinants are examined. As there are 23 proximate determinants, 23 equations are estimated in the first-stage. These results tell us how much a change in a background factor will influence a proximate determinant. The impacts of maternal education on the proximate determinants are shown in Table 5.9. These reveal the hypothetical average

"choices" of health inputs made by women with different levels of schooling -- controlling for all other socio-economic factors.

The second stage estimation is the hybrid equation estimated above. This tells us the impact of a proximate determinant on mortality. The impact of a background variable through particular channels is simply the product of the estimated behavioral coefficient linking education with a particular behavior and the estimated biological coefficient measuring the effect of that behavior on mortality. The sum of all these pathways yields an estimate of the total effect of maternal education mediated by the proximate determinants. These multiplicative sums are presented in Table 5.10 for the neonatal period and Table 5.11 for the postneonatal period.

Most of the effect of maternal education on neonatal mortality is mediated by the proximate determinants. That is, children of lower educated women have higher neonatal mortality because of different "choices" of health inputs in the child production function. The inputs which are most important can be examined by estimating their relative contributions to an observed differential. For example, children of women with no schooling face 1.7 times the risk of high educated women (7+ years) due to different health inputs. Which "inputs" are most important?

In Figure 5.11, I compare the differences in inputs "chosen" by mothers with no schooling and mothers with 9 or more years of schooling. The largest contribution is the absence of a toilet which increases the relative odds of neonatal mortality by 1.3 times. This is followed by the lack of tetanus immunization and delivery assistance which combine to increase the relative odds by 1.2 times. Small size at birth increases the risk another 1.1 times. The greater tendencies of mothers with no education to have children at high parities (6+) is offset by their tendency to have children at higher ages (40-45). On net,

the deleterious effect of high parity is offset by the apparent beneficial effects of childbearing in the 40-45 age range to yield a risk .97 times that of higher educated mothers. These effect of these choices multiply to yield the observed differential of 1.7 ( = 1.3 x 1.2 x 1.1 x .97) = 1.7. The other pathways yield small effects which have a multiplicative sum of 1.

Although short birth intervals have a strong effect on infant mortality, there are not large differences in birth intervals by maternal education (see Table 5.9). Hence, mortality differences between high and low educated women cannot be attributed to differences in birth intervals, nor reproductive patterns in general. This is consistent with previous studies of the effect of maternal education on infant mortality (Barbieri (1989) and Bicego and Boerma(1991)). However, these studies failed to examine interaction effects. That is, short birth-intervals may be especially deleterious to lower educated women. I attempted to interact education with birth-spacing but found no statistically discernible effects . This is most likely due to small sample sizes. So, this remains an open research question.

Figure 5.12 displays the direct and indirect impact of maternal education on postneonatal mortality. While there are very large educational differences in postneonatal mortality, very little of this is mediated by the proximate determinants. Children of mothers with no schooling face 1.18 times the risk of children of higher educated mothers due to different inputs in the health production function. The chief inputs responsible are: lack of delivery assistance, absence of toilets, and birth parity and age at birth. Lack of delivery assistance at birth elevates postneonatal risks by a factor of 1.15. The absence of toilets raises the risk by 1.11. Again, there are off-setting reproductive factors, the deleterious effects of high parity and the beneficial effects of high maternal age. The net effect is lowered risk among children of women with no schooling-- who face .93 times the risk of higher

educated mothers. These multiplicative effects yield a combined effect of 1.18. The other pathways have small effects which result in a multiplicative sum of nearly 1.

Examination of child mortality by cause provides a better basis for causal explanations of the correlations between risk factors and child death. In particular, we are interested in a deeper understanding of the direct and indirect effects of maternal education on child mortality. The main drawback is that we divide our sample into smaller units which means it becomes more difficult to observe statistically significant relationships in the data.

Determining cause of child death is a difficult task in developing countries. In the Bolivian DHS, mothers were asked what type of accident or illness was responsible for their child's death. The mothers' original responses to these questions about the types of illnesses and accidents have been recoded by the DHS according to the International Disease Classification Codes. Use of such reports may be criticized on the grounds that a mother may have little idea of the medical cause of her child's death or that native understanding of disease classification may differ substantially from the international disease classification system. (E.g., How does one classify maternal reports that a child died from susto, literally "fright"?) In my opinion, however, this argument underestimates the extent to which public health education campaigns have succeeded in spreading knowledge of modern disease terminology.

In addition, the mothers were asked a series of questions about symptoms prior to a death (a "verbal autopsy"). From this list of symptoms DHS developed a disease algorithm to translate from these responses into probable cause of death. There is no way of assessing the accuracy of either of the mothers' reporting or the DHS symptom algorithm. In this study, the mother's own reporting is used rather than the DHS algorithm, since it seems best not to introduce another level of uncertainty in this analysis.

In the neonatal period, three sets of causes are examined: non-infectious, infectious, and unknown. Non-infectious causes are dominated by reports of trauma at birth, asphyxia, hypoxic stress, and tetanus. These deaths occur in the first hours or days of life. The second set of causes examined are infectious diseases consisting of diarrheal and respiratory deaths. The third set consists of deaths whose causes were unknown to the mother or could not be classified.

In the post-neonatal period, three sets of causes are examined: infectious diseases spread by fecal-oral mode of transmission, airborne infectious diseases, and all other causes. Infectious diseases are divided by the two main modes of transmission since health intervention efforts are expected to have different impacts on these diseases. For example, sanitation, washing hands, and better food preparation techniques would be expected to reduce fecal-oral transmitted diseases; whereas reductions in crowding should have major impacts on airborne diseases.

The multinomial logit is used to model discrete outcomes which can take on more than two values. In this case, four outcomes are modeled: life, death from cause A, death from cause B, and death from cause C. The multinomial logit has properties similar to that of the binomial logit model, except that in the case of the binomial, one probability is estimated for each individual, whereas in the multinomial model m-1 probabilities are estimated for each individual, where m is the total number of choices.

An additional property of some types of multinomial models is the "Independence of Irrelevant Alternatives" (IIA) property. When the IIA property holds, the odds ratio for any two alternatives is independent of any other alternatives in the choice set. This property is undesirable in situations in which alternatives are believed to be dependent on one another. A frequently cited example is the red bus/blue bus problem. Assume that individuals are indifferent between taking a car or taking a bus. If an individual is faced with the choice between taking a car or taking a red bus, the predicted probability of taking the red bus should be 1/2. If the individual is faced with three choices: car, red bus, and blue bus, we assume that the individual views red and blue buses similarly. That is, these choices are not independent of one another. The expected outcome is car (1/2), red bus (1/4), and blue bus (1/4). But, a multinomial model with the IIA property assumes that all alternatives are independent. Hence, it predicts the new transportation probabilities as car (1/3), blue bus (1/3), and red bus (1/3).

There are two ways of incorporating dependence of alternatives in a multinomial model. The first is estimation of the multinomial model with alternative-specific constants. These intercepts force the sums of the predicted probabilities over all individuals for each alternative to equal the observed frequency of each alternative. Thus, if individuals do in fact view red buses and blue buses similarly and this is reflected in their observed choices, the logit equation will faithfully reflect this dependence. The second is to estimate a nested multinomial logit model. Those alternatives which are deemed to be similar are grouped together into nests. The IIA property then holds within but not between nestings. In the red bus/ blue bus example, the nested logit views the choice of transportation as a choice between cars and buses and a choice between red and blue buses. The bus alternatives form a nesting. The latter method is preferable since it is more flexible and can yield estimates of the degree of dependence among nested factors. The analysis reported in this chapter uses the multinomial logit with alternative-specific constants in order to mitigate the effects of violation of the IIA property. Implausible estimation results were obtained from a nested logit regression, probably due to severe multicollinearity problems owing to both small sample sizes and use of categorical independent variables.

In estimating the multinomial logit, it is sometimes the case that the relationship between an outcome category (death from cause A) and a risk category (e.g., twin) becomes completely predicted (either 0 or 1) for some set of observations. That is, no twins die from cause A. In this case, the parameter estimate is positive or negative infinity and estimation is not possible. The model can be estimated by either re-defining the risk category or dropping the variable from the model. In the case of neonatal mortality, the variable twin had to be eliminated.

The results from the multinomial logistic analysis are presented in Tables 5.12. and 5.13. These equations are estimated to help us understand the indirect and direct impacts of maternal education on child mortality.

The binomial logit analysis indicated that most of the effect of maternal education on neonatal mortality was mediated by the proximate determinants. Specifically, there were four important pathways: absence of toilet, lack of tetanus immunization, lack of medical assistance at delivery, and small size at birth. Are these pathways measuring the direct effects of these behaviors on mortality? Or are these "proximate determinants" actually proxies for other unobserved factors? These are crucial questions to answer, since the

usefulness of the proximate determinants analysis lies in its promise to reveal the bio-medical pathways through which socio-economic factors influence child mortality.

The presence of a toilet has a strong effect on reducing risks of death from infectious diseases. Those with a sewer line hookup or a septic tank face a relative risk of death from infectious disease 0.17 times that of those lacking such facilities. This would tend to indicate that presence of a toilet is directly measuring a health effect rather than proxying for wealth. However, strong effects are also seen in reducing risks of death from non-infectious diseases, though the large standard error associated with this estimate make it statistically insignificant from zero.

Tetanus immunizations seem to be a proxy for other factors, which was noted when it was observed that tetanus immunizations had larger impacts in the postneonatal period than in the neonatal period -- where tetanus deaths occur. Further evidence is seen in examining the cause of death in the neonatal period. Children whose mothers received tetanus immunizations were at reduced risk from both infectious and non-infectious diseases.

Delivery assistance appears to raise the risk of death from infectious disease and lowers the risk from non-infectious disease. No plausible explanation for these estimates is found-- and delivery assistance is presumed to be proxying for some other factor.

The final pathway which accounted for most of the educational differentials in neonatal mortality was size at birth. One would expect differences in birth weight based on economic characteristics rather than education. The most plausible explanation is that, despite attempts to control for income, maternal education continues to partially proxy for income effects. Children of women with no education were more likely than children of higher educated women (9+ years) to be judged to be "much smaller than normal" at birth. These children are found to be at increased risk of death from both infectious and non-infectious diseases.

Based on this analysis, many of the proximate determinants appear to be proxies for other unobserved behaviors. The proximate determinant equation appeared to explain most of the differences in neonatal mortality between children of mothers with no schooling and those with higher educated mothers. However, examination of the beneficial effects of these bio-medical pathways revealed effects that are unlikely to have been caused by some of these factors. This casts doubt on the ability of proximate-determinant analyses to yield useful policy insights.

In the postneonatal period , the proximate determinants explained relatively little of the effect of education on mortality. Here, we focus attention on the direct and unmediated effects of maternal education on mortality by cause and a clearer picture emerges. Maternal education is strongly associated with reduced risk from fecal-oral diseases. By contrast, no statistically significant effects are evident in risk of death from airborne diseases. In fact, these parameter estimates (with large standard errors) indicate that children of higher educated women (9+ years) face higher risks of death from airborne diseases.

This result can be derived in a considerably simpler fashion by a cross tabulation of postneonatal mortality by cause of death and maternal education. Figure 5.13 shows that there are virtually no educational differences in mortality rates from airborne diseases nor from non-infectious causes. The educational differential observed in postneonatal mortality is solely a reflection of different rates of mortality from fecal-oral diseases such as diarrhea.

What can we conclude from this? It may be that mothers with little or no schooling were more likely to say "diarrhea" when asked what their children had died from, so this evidence might be dismissed as reporting error. This may, on the other hand, be important evidence of the beneficial child care practices followed by higher educated mothers. There is not much a mother can do to prevent infectious disease spread by air. However, there are many health behaviors such as hand washing after defecating, boiling water, proper food preparation, and sanitary disposal of diapers -- which can prevent fecal-oral spread diseases. Similarly, there are not many non-medical things a mother can do to prevent child death from airborne diseases such as measles, TB, or pneumonia. Such diseases require medical interventions in clinics or hospitals. However, there is much a mother can do such as continued breast feeding or use of mates or oral rehydration salts -- to prevent death from fecal-oral diseases such as diarrhea. Hence, the fact that maternal education is strongly associated with reduction of the risk of death from fecal-oral diseases and unrelated to death from airborne diseases may be a reflection of the adoption of these non-medical and highly effective preventative and curative behaviors.

This analysis attempted to assess the impact of maternal education on mortality net of wealth and income effects. Maternal education was found to exert a strong effect on mortality risks in both the neonatal and postneonatal period independent of husband's education, husband's occupation, and family and geographic characteristics. There does appear to be a micro-level relationship between maternal education and child survival which underlies Bolivia's mortality decline. However, this analysis was unable to determine the bio-medical linkage between maternal education and child survival. A causal explanation of this relationship remains unproven.

A proximate determinants framework was used to attempt to assess the pathways by which maternal education influences mortality. Application revealed that most of the effect of education on neonatal mortality appeared to be mediated by these proximate determinants. Specifically, the results indicate that lower educated mothers have higher rates of neonatal mortality because they tend to live in homes without sanitiation facilities, tend not to receive tetanus immunizations nor assistance at delivery and tend to deliver low birthweight babies. However, close examination of the impact of these bio-medical factors on specific causes of death indicated that most were proxying for some unmeasured behaviors.

In the postneonatal analysis, the proximate determinants appeared not to mediate the effects of maternal education. It was found that most of educational differentials in mortality were attributable to fecal-oral diseases. This suggest that higher educated mothers may have adopted preventive and curative domestic child care practices which greatly reduce their children's risks of death from fecal-oral diseases. Thus, while the specific bio-medical mechanism remains unidentified, maternal education appears to influence child survival through changes in behavior.

A key implication of this research is the need for the collection of more appropriate data in health interview surveys. In the DHS-I Surveys, the three questions asked of mothers with respect to all children (living and dead) pertain to prenatal and natal medical care -- care which is likely to affect neonatal mortality. Yet, most child deaths occur in the postneonatal period and most are preventable through adoption of non-medical behaviors. Health interviews ought to be restructured to focus on these types of behaviors.