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

The term "stunting" was introduced by Waterlow over a decade ago to "describe what one actually sees: a deficit in attained length or height of children compared with international standards." This purely descriptive term also "conveys the impression that the child does not have to be short, that something has gone wrong." This chapter examines stunting among Bolivian children, who are among the shortest children in the world. Has something gone wrong? Is their short stature relative to an international standard a sign of abnormality or is comparison to an international standard inappropriate? Are these children simply "naturally" short? Is this a reflection of the genetic adaptation of a population evolving in a high altitude environment like the Andes?

Most of the previous work on child growth in Bolivia has been conducted in the Andes by physical anthropologists. (See accompanying table). The seminal study in the field was Frisancho and Baker's 1970 study of rural Quechua living in Nuñoa in the Peruvian Andes. The authors observed through anthropometric measures that this population was growing much more slowly compared to international growth standards, yet dietary surveys indicated that they were sufficiently well-nourished. While mindful of the potential role played by malnutrition in slowing growth, the authors cautiously concluded that hypoxia (oxygen shortage) and/or genetic adaptation to hypoxia are more likely explanations of the slower growth. They wrote:

"These findings suggest that factors other than malnutrition or caloric deficiency, such as high altitude hypoxia and/or genetics are influential in the slow and prolonged growth of Nuñoa Quechua population. This conclusion by no means implies that all growth variation reflected in slow and prolonged development of high altitude populations is solely related to the effects of hypoxia. Any generalizations about growth at high altitude must be done with great caution, and until extensive field studies are done, the role of nutrition cannot be disregarded."

Subsequent studies confirmed the existence of a pure hypoxia effect on growth retardation by comparing growth of middle-class children of European ancestry at high and low altitudes [Stinson (1982); Schuttee (1983); Greksa (1990)]. However, the hypoxia effect was found to be rather small -- not nearly large enough to account for differences observed between villages, between social classes, and between ethnic and racial groups.

The most oft-cited factor to account for these observed differences in growth is chronic malnutrition. Greksa (1984, 1986) suggested this as the most probable factor accounting for the better growth of Aymara children in La Paz when compared to Quechua children in Nuñoa. However, he notes that: "... although nutritional variation is a likely cause of the different patterns of growth observed in Nuñoa and La Paz, altitude and genetics (or interactions between any of these factors) cannot be eliminated as possible causal agents." Leonard (1989) revisited the original Nuñoa community studied by Frisancho and Baker. He demonstrates that today children there are growing faster than two decades ago. He asserts that this secular trend is unlikely to have been caused by genetic changes and presumably is the result of nutritional improvement. More evidence on the role played by malnutrition is provided by de Meer (1993), who observes large variation between communities which are situated at similar altitudes. de Meer ascribes these growth differentials to differences in the agricultural settings among these communities and their likely impact on food security and availability.

The impetus for studying child growth in the Andes came from the desire to understand "how populations living with a simple technology had biologically adapted to their environments." The original research centered around high-altitude hypoxia and genetic adaptation as explanations of slow child growth in the Andes. The general consensus of the anthropological literature that emerged over the next two decades is that chronic malnutrition rather than high-altitude hypoxia or genetic adaptation is responsible for the observed growth retardation. However, from the start all three factors -- high altitude hypoxia, genetic adaptation, and malnutrition -- have been considered as plausible causal factors and none has been conclusively ruled out.

After two decades of research, we are still uncertain of the contributions of high altitude hypoxia, genetic adaptation, and malnutrition to the slow growth of children for two main reasons. First, studies of child growth in the Andes conducted by physical anthropologists have examined a small number of children in a total of ten different elevations. It is difficult to determine the impact of altitude when only two or three different locations are being compared. Second, most indigenous populations live at high altitudes and are poor. It is difficult to separate the effects of genetic adaptation, high altitude hypoxia, and malnutrition on child growth.

The data set used here is better able to illuminate these problems. The DHS collected height and weight measurements of children aged 3 - 36 months in 1989. By merging the DHS data set with data collected in the 1988 National Population and Housing Survey, the villages, towns, and cities in which the DHS survey was conducted were identified. Altitude data for these places was collected based upon a topographic map of Bolivia and the Atlas Censal de Bolivia (1982). The data set thus created contains anthropometric information on 2,612 children in 160 different locations. To my knowledge, it is the largest data set linking altitude information with anthropometric measurement of children in a developing country setting.

The height measures of children (and length measures of infants) can be standardized by age and sex by converting the measured height into a measure of deviation from a particular standard age and sex pattern of growth. The age and sex standardized values are known as "Z-scores". The formula for their calculation is:

Equation 4.1

There are three reasons for using Z-scores as the unit of analysis. First, rather than examining boys and girls and each age separately, one can combine these observations together into one large sample for analysis. Second, if a common standard is used, comparisons between different studies can be made. Third, and most importantly, if this common standard is assumed to be a norm, then large negative Z-scores are an indicator of abnormally slow child growth.

The World Health Organization recommends using the U.S. National Institute of Health standard growth curve of children. This standard was derived from two distinct sources. The first set of data was collected by the Fels Research Institute as part of a longitudinal study of growth. Information was collected on 867 children at ages 1, 3, 6, 9, 12, 24, 30, and 36 months from white, middle class families living near Yellow Springs, Ohio. The second source was collected by the National Center for Health Statistics (NCHS) in a series of health surveys on a representative sample of U.S. children (in 1963-65 for ages 6-11, in 1966-70 for ages 12-17, and in 1971-1974 for ages 2-17).

Several problems in the growth standard have been noted (Dibley, et. al. 1987). Most importantly, a discontinuity has been observed in the standard curve at age 24 months where the data from Fels is splined to the data from NCHS. Children from the Fels group are somewhat taller and heavier than children from the NCHS group. Hence, estimates of low height-for-age will be higher when children are compared against the Fels standard during the first 24 months of life, and will then drop abruptly when these children are compared against the NCHS standard. It appears likely that the observed worldwide pattern of wasting and stunting (a peak in the ages 12-24 months) is partly due to this inconsistency in the standard. To avoid this problem, Dibley, et. al. recommend separate tabulation of data for children under and over 24 months.

Before presenting the analysis and results, three potential data problems are discussed. These are age misstatement, censoring of observations due to mortality, and measurement error associated with our proxy for racial groups. The latter is shown to be the most serious problem.

There are two sources of measurement error in determining a Z-score: the measured height of the child and the stated age. The three figures on the following page present the results of a simulation which examines the effects of these two types of measurement error on child Z-scores. Figure 4.1 shows the Z-scores for 10,000 children of exact age 22 months and exact height of 81 cm. All 10,000 children have the same height-for-age Z-score of -1.50. Figure 4.2 demonstrates the effect of error in height measurement -- characterized by zero mean and a standard deviation of 0.34 cm. This estimated error is taken from an analysis of measurement variability in an anthropometric study of pre-school children in Guatemala. Figure 4.3 demonstrates the effect of an error in age mis-statement -- characterized by a mean of 0.8 months (age is over-stated) and a standard deviation of 3 months. This estimate is taken from a study of Bangladeshi pre-school children. Comparison of Figures 4.2 and 4.3 reveal that the main source of error in

measurement of Z-scores in developing countries is likely to be age misstatement.

Bairagi, et. al. (1987) examine the effects of age misstatement on weight-for-age and height-for-age measures using data collected on 679 Bangladeshi children aged 22 to 59 months. Information on the children's true ages were taken from vital registration data and compared against the ages reported to the interviewers when the child was measured. On average, the ages of these children were overstated by 2 - 3 months. The standard deviation around this mean was between 7 and 10 months. Thus, there was a considerable amount of understating of age as well. In addition, the authors noted that child ages were not misstated at random. First, children who were big for their age tended to be reported as older, and small-for-age children as younger. Second, uneducated mothers were more likely to misstate the age of their children. This differential reporting resulted in an artifactual relationship between mother's education and height-for-age and weight-for-age measures.

In DHS surveys, the ages of children are derived from maternal reports of birth dates rather than from maternal responses to questions about a child's current age or an interviewer's best guess. Hence, the type of age misstatement present in Bairagi, et. al.'s study is unlikely to be present in DHS surveys. I have no way of assessing general age mis-reporting, but we do not observe significant age heaping. (See Figure 4.4). However, a peculiarity in the age distribution worth mentioning is the lack of 3 and 36 month aged children. As seen in the Table 4.2 below, 50% of 3-month-olds and 63% of 36-month-olds were not measured in the survey. This suggests that interviewers may have been uncertain about whether or not to measure these children. Thus, this peculiar age distribution may reflect interviewer confusion over whom to measure rather than age misstatement by the mothers.

While noting the need for further research on this subject, for the purposes of this analysis -- whatever the effect of age mis-statement on estimates of Z-scores -- the assumption will be that such mis-statement is minimal in Bolivia and hence of minimal import to this analysis.

In the Bolivian DHS, measurements were taken on 2,612 children who were born in the three years prior to the survey -- an additional 560 children were not observed due to censoring from mortality. Hence, cross-sectional information on malnutrition may be unrepresentative of the nutritional status for a birth cohort of children. One could imagine

that the more malnourished children died prior to observation. In a recent assessment of the bias introduced by mortality, Boerma, Sommerfelt, and Bicego (1992) found that these effects were likely to be small and generally would not affect analysis of the determinants of malnutrition.

Age misstatement is a special case of the more general problem of response error. Respondents can misstate their characteristics (such as education and their age) and their health behaviors. In the Bolivian DHS, it appears that bilingual respondents overwhelmingly reported Spanish as the usual language spoken at home. I believe this to be an example of mis-reporting.

There are two reasons why this mis-reporting is likely to have occurred. First is the desire of respondents to appear "modern" (i.e., non-indigenous) to the interviewer. Though the rise of indigenous political movements in the 1980's (e.g., the Katarista movement) may indicate the beginnings of a profound change in social consciousness, indigenous culture in Bolivia has generally been afforded a low social status.

Another factor could be the desire on the part of the interviewer to justify use of the Spanish language questionnaire. There were significant problems in implementing the Quechua and Aymara language versions of the questionnaire. The main problem was that the Quechua and Aymara versions (translated from Spanish) proved to be of the "high-form" of the language -- generally unintelligible to those used to speaking the "common-form" of the language. Hence, most Quechua and Aymara, irrespective of their level of Spanish proficiency, were given a Spanish language questionnaire. The bilingual interviewer would then spontaneously translate the questions as needed.

Shown in Figure 4.5 are the responses given in the DHS to the question concerning the language usually spoken at home. "Spanish" is the overwhelming response. For those who stated that they spoke an indigenous language, ability to speak Spanish was also measured. Unfortunately, all that is known about the Spanish speakers is that they speak Spanish, so that monolingual and bilingual Spanish speakers cannot be distinguished.

Information collected a year earlier as part of the National Population and Housing Survey presents a very different picture. In Figure 4.6, the responses of women aged 15-49 to the question: "Which languages do you know how to speak?" have been tabulated. Over half the respondents reported speaking an indigenous language.

In comparing the two surveys, the reported percentages speaking Quechua Only or Aymara Only are similar. This is strong evidence against the hypothesis that interviewers, rather than interviewing their pre-assigned households, sought out those households in which Spanish was spoken. Rather, it appears that large numbers of bilingual Quechua and bilingual Aymara classified themselves as living in households in which the majority of people usually speak Spanish.

Whether they are mis-classifying themselves is not important. Though I do not believe this to generally be the case, one could imagine that bilingual Quechua and Aymara women really do live in primarily Spanish-speaking households -- perhaps they married a Spanish-speaking husband, or perhaps their children increasingly use Spanish at home as

well as at school. What is important is that I cannot tell the difference between bilingual and monolingual Spanish speakers who report themselves as living in primarily Spanish-speaking households. This is an important problem for my study, as I want to control for genetic background in examining child growth. I will use the response to the language question as my primary proxy for this. It now appears that this proxy is measured with a considerable amount of error.

The use of stunting as a measure of malnutrition in the presence of possible genetic differences in child growth and differences in exposure to the growth-slowing effect of high-altitude hypoxia will now be examined. In this section, the Z-scores of Bolivian children are analyzed with respect to characteristics of the child (age, sex); characteristics of the mother (education, language spoken); characteristics of the place of residence (rural/urban, altiplano/highland valleys/tropics).

The mean Z-scores for various population groups in Bolivia are presented in Figures 4.7, 4.8, and 4.9. First, it should be noted that all the values are negative. These groups are all growing at rates significantly slower than the international standard. Next, consider growth differences based upon the age and sex of the child (See Figure 4.7). Z-scores become progressively worse with age. This is expected as stunting is a cumulative process. There are no apparent sex differences. This is indicative of similar child care and feeding practices for both boys and girls.

In considering growth differences based upon characteristics of the mother (Figure 4.8), there are large differentials based upon education and language spoken. Education, also

known to be strongly associated with infant mortality, is similarly strongly associated with child growth. The reasons for such a correlation may be related to differences in knowledge, power, and/or wealth among women with different levels of education. There are two very different explanations for the observed correlation between speaking an indigenous language and slow child growth. The first is that these children are suffering from chronic malnutrition -- because speaking an indigenous language in Bolivia is associated with ignorance, powerlessness, and/or poverty, all of which contribute to a lack of resources to provide for healthy child growth. The second is that these children grow more slowly because they are genetically different from the population to which they are being compared. These competing explanations cannot be resolved by examining cross-tabulations.

In considering growth differences based upon place of residence (Figure 4.9), one notes large differentials between urban and rural areas and among the three ecological regions of Bolivia. Child growth is significantly slower in rural areas. Multiple explanations of this phenomenon are possible. There are more indigenous speakers in rural areas, there are low levels of education in rural areas, and people in rural areas are often quite poor and generally have little access to modern health care. The strong differentials observed among regions may be due to differences in diet, in access to health care, in standards of living, in infectious disease environments, or to hypoxia at the higher altitudes present in the Altiplano and Highland Valleys.

A multivariate linear regression equation is used to attempt to predict Z-scores based upon characteristics of the child (age, sex), characteristics of the mother (educational attainment, language usually spoken, ability to speak Spanish), and characteristics of the place of residence (urban/rural; Altiplano/Highland Valleys/Tropics, Altitude). These coefficients are graphed in Figure 4.10 and the regression is presented in Table 4.3.

These results argue against a genetic interpretation of slow growth among Bolivian children. The small coefficient on monolingual Quechua speakers (coeff =

-0.34) can be compared to the large differentials seen in the cross-tabulations: children of monolingual Quechua mothers were on average 1 standard deviation below Spanish children. By contrast, the multivariate regression reduces this difference by 2/3 once other characteristics of the mother, child, and place of residence are considered.

There is significantly slower child growth among families in which the mother speaks only Aymara (coeff = -.90). This should not be considered important evidence in support of genetic differences in growth between Bolivian children and the international standard for two reasons. First, there appears to be no significant effect on child growth (coeff = 0.02) for children of mothers who speak Aymara and Spanish. The slower growth of monolingual Aymara compared to bilingual Aymara is not likely to be due to due to genetic differences between these populations. More plausible is that this growth difference is due to socio-economic differences associated with the ability to speak Spanish. Second, even if this slow growth among this group were solely attributable to a genetic difference, this could not be the major reason for the observed slow growth of the majority of Bolivian children.

Further evidence of the minimal effect of genetic differences on slowed child growth is seen in examining bilingual speakers. As noted previously, families in which the mother is a bilingual Aymara speaker do not have significantly slower child growth (coeff = 0.02). Also, families in which the mother is a bilingual Quechua speaker show only slightly slower child growth (coeff = -0.21). As noted in Section 3.3, both of these variables are measured with considerable error. This casts doubt as to the true magnitude of these parameters.

This measurement problem is analogous to the Errors in Variables problem in econometrics. In classic OLS regression, only the dependent variable is assumed to be measured with error. OLS regression becomes biased and inconsistent when independent variables are measured with error.

There are three main approaches to this problem. The most common practice is to ignore it, assuming the measurement errors to be minimal. The second approach involves the instrumental variable technique, in which the poorly measured independent variable is replaced with a variable which is correlated with that variable but uncorrelated with the measurement error. The other approach involves use of the variance of the measurement errors. This is rarely used because the error in variables is usually unknown.

In attempting to correct for this problem, two of the previously mentioned approaches will be combined: instrumental variable and knowledge of the measurement error variance. In this case, variables measured with error in which the error variances are unknown are replaced with variables for which the error variances can be estimated. Specifically, the discrete variable: "Do you speak Quechua?" is replaced with a continuous variable reflecting "Quechua-ness." This variable takes on a value between 0 (reflecting no degree of Quechua-ness) to 1 (reflecting 100% Quechua-ness). "Quechua-ness" is defined as the probability that the respondent speaks Quechua. This probability is derived in the following manner. The 1988 National Population and Housing Survey is assumed to correctly report the percentage of Quechua and Aymara speakers in the village. The difference between this percentage and that reported in the DHS is assumed to be due to Aymara and Quechua speakers reporting themselves as Spanish speakers. The Quechua and Aymara speakers are given a value of 1 in Quechua-ness and Aymara-ness, respectively. The Spanish speakers in the village are given some fractional value of Quechua-ness and Aymara-ness such that the sum of Quechua-ness and Aymara-ness for a village will reflect the percentage derived from the ENPV.

An example is presented in Table 4.4 below. Assume that in village X in 1988 50% (100 of 200) of the respondents to the ENPV Survey classified themselves as Quechua speakers. A year later in the DHS, 19 classified themselves as primarily Spanish speakers and 1 as primarily a Quechua speaker. This yields a value of 5%. The 1 Quechua speaker receives a value of 1.0 for Quechua-ness. In order to match the ENPV percentage of 50%, it is assumed that 9 of the 19 Spanish speakers are bilingual speakers. Hence, all 19 Spanish speakers are assigned a value of 0.473 (9/19) for Quechua-ness.

The table below compares the responses given in DHS to estimates of Quechua-ness and Aymara-ness. According to the original DHS responses, 1261 people or 22% of the sample classified themselves as primarily non-Spanish speakers. Using the technique outlined above, all the primarily Spanish speakers were reassigned as partial non-Spanish speakers. Thus, some are classified as 10% Quechua, others as 15%, etc. The sum of these fractional value yields an estimate of 1,078 "equivalent Quechuas." That is, I estimate that there are 1,917 Quechua-speakers in the sample -- 839 of whom reported themselves as Speaking Quechua and 1,078 of whom identified themselves as speaking Spanish. These 1,078 people cannot be precisely identified. However, who they probably are can be determined in the manner outlined above.

These measures of Quechua-ness and Aymara-ness have two useful properties: a mean error of zero and an easily calculated error variance. The table below evaluates the values of the variance in the true but unobserved X and the error of our instrumental variables.

With this information, the bias in OLS estimation caused by this measurement error can be corrected. Theil (1961) derived approximate formulas for the bias when two variables are measured with error. The formulas are:

Equations 4.2

Table 4.8 on the following page presents two regressions: the first with the incorrectly measured variables Quechua and Aymara, the second with the incorrectly measured variables Quechua-ness and Aymara-ness. Though the parameter estimates are biased in both regressions, in the latter one we can estimate the direction and magnitude of this error. Hence, we can arrive at an unbiased estimate of the effect of speaking Quechua or Aymara on child growth. Application of Theil's approximation as reported in equation 4.2 yields the following results.

The correction shows that the true effects are still rather modest -- on the order of 1/2 and 1/4 of a standard deviation below the international growth standard. Therefore, it is exceedingly unlikely that the growth difference observed between Bolivian children and the international standard (of about 1½ standard deviations) is principally due to genetic differences between Bolivian and United States children.

This study contributes to the accumulating evidence that well-nourished, healthy children throughout the world grow at similar rates from birth until the teenage growth spurt. Hence, there is no strong evidence that the international growth standard needs to be adjusted to reflect genetic differences between racial or ethnic groups in Bolivia.

Altitude seems to have an important impact on child growth. Indeed, the regional differences in Bolivia previously observed in the bivariate analysis are largely explained by altitude. In the bivariate analysis, children on the Altiplano were on average 0.77 standard deviations below children living in the tropics. By contrast, in the multivariate analysis which controls for altitude, this regional effect is reduced by 2/3 to a more modest, -0.23. Similarly, the bivariate analysis showed children living in the highland valleys to be 0.44 standard deviations below children living in the tropics. In the multivariate analysis, the effect barely discernible at -0.05. The most common explanation for the effect of altitude on child growth is slowed growth due to oxygen shortage (hypoxia).

Recently, Yip, et.al. (1988) examined this relationship among U.S. children and found a significant altitude effect. Shown in Figure 4.11 are the mean Z-scores by elevation for children receiving medical care under the Medicaid program in the western mountain states of the United States taken from Yip (1988). The Z-scores of Bolivian children follow a similar slope. Statistical comparison of these two slopes indicates that one cannot reject the null hypothesis that there is no difference between the altitude effect observed among U.S. children and that observed among Bolivian children. Since a similar altitude effect is being observed, this suggests the effect of pure high altitude hypoxia rather than an effect of lower access to health care at higher altitudes.

Further evidence of this hypoxia effect is seen in Figure 4.12. This graphs the Z-scores of children by educational category and elevation. An altitude effect is seen among all educational groups.

These data also cover a higher range of elevation (over 2,500 meters) than those studied in the U.S. case. Here, it can also be demonstrated that a similar growth-slowing effect is present at extreme altitudes. An F-test fails to reject the null hypothesis that there is no difference in the altitude effect when comparing elevations below 2,500 meters to elevations above 2,500 meters. Hence, it appears that the effect of hypoxia on slow child growth exhibits a linear dosage effect.

The strongest correlate of slow child growth in Bolivia is the educational attainment of the mother. These differences are unlikely to be due to genetic or altitude differences. Rather, they reflect the higher levels of chronic malnutrition in families with low levels of educational attainment. This may be due to differences in knowledge, wealth, and power among these families. Such differences are not readily identifiable in large-scale surveys such as the DHS.

Children who are chronically malnourished grow more slowly than well-nourished children. Hence, stunting is commonly used as a proxy for assessing chronic malnutrition in a population. But other factors -- unrelated to ill-health -- have been proposed as contributing causes of stunting in a population. In this chapter, two such factors have been considered: genetic differences between population groups (Aymara, Quechua, and European/Mestizo descent) and high-altitude hypoxia.

It does not appear that genetic differences in Bolivia are of a magnitude to affect assessment of chronic malnutrition. Hence, the use of one international standard without respect to presumed genetic differences in populations appears justified, based on these data. In contrast, the effect of altitude appears to be rather significant. This effect is large enough to warrant reconsideration of a single international growth standard for children throughout the world.

The percentages of children stunted before and after adjusting for the growth-slowing effect of hypoxia are presented in Figure 4.13. For the country as a whole, one third (35%) of children can be classified as abnormally low height for age, being more than two standard deviations below the international reference standard. Some of this slow growth is attributable to hypoxia and not chronic malnutrition. But, even after controlling for this hypoxia effect, one fourth of Bolivian children would still be classified as chronically malnourished.

The main problem in failing to account for the hypoxia effect is in determination of the distribution of malnourished children among regions in Bolivia. Use of one uniform standard would result in the erroneous conclusion that chronic malnutrition is heavily concentrated in the Andean regions of Bolivia. Estimates based on Z-scores would indicate that half of all malnourished children live on the Altiplano. (See Figure 4.14). Z-scores of children are lower on the Altiplano, but this is due in large measure to the effect of high-altitude hypoxia and not malnutrition. As revealed by altitude-adjusted

Z-scores (Figure 4.15), malnutrition is widespread throughout Bolivia in the highland valleys and lowlands as well as the Altiplano.