Stunting is widespread among children living in environments characterized by poverty, poor nutrition and a high prevalence of infectious disease. Stunting is an important public health problem in developing countries because of its association during childhood with poor functional outcomes such as impaired cognitive development (Pollitt et al. 1995 and Brown and Pollitt 1996), increased susceptibility to infection (Tomkins 1988) and increased risk of mortality (Bairagi et al. 1985, Chen et al. 1980). Long-term consequences of childhood stunting include short stature, reduced work capacity (Spurr 1988) and elevated risk of poor reproductive outcomes. Recent research also suggests that stunting in childhood may increase later risk of obesity and chronic diseases (Barker 1994).

Most stunting occurs in the first 2-3 y of life. Although there may be some biological potential for catch-up growth through adolescence, especially when development is delayed, few opportunities for catch-up exist among children who remain in the poor environments that caused them to become stunted in the first place. Thus, once children have become stunted, they are likely to remain stunted into adulthood (Martorell et al. 1994). It is therefore imperative to understand the causes of stunting early in infancy and childhood, so that preventive measures can be taken.

Waterlow (1994) and Allen and Uauy (1994) emphasized that despite numerous studies, relatively little is known about the causes and mechanisms of stunting. There are strong associations between poverty and stunting; however, relatively few longitudinal studies have identified more proximate causal factors related to poverty. Furthermore, effects of such determinants at different ages have not been identified. The weaning period of 6-12 mo is assumed to be a particularly vulnerable time, but growth retardation continues among toddlers as well. Previous studies have examined the prevalence of stunting at different ages, but typically do not account for the fact that a child who is stunted at 24 mo of age, for example, was likely to have become stunted at a much earlier age. Thus age-specific etiology is not identified.

We used prospective data collected bimonthly on a large sample of children in Cebu, Philippines, to study stunting in the first 2 y of life. The highly detailed longitudinal data allowed us to make new and valuable contributions to our understanding of stunting through the following approaches: 1) examination of the incidence rather than prevalence of stunting, 2) modeling of the effects of age-specific proximate determinants of stunting, and 3) the use of longitudinal, multivariate statistical methods that correct for biases inherent in much of the previous work on stunting.

SUBJECTS AND METHODS

In the hazard framework, once a child becomes stunted, he or she is censored, and subsequent recovery and relapse are ignored.⁶ We estimated the log odds that a child would become stunted during each 2-mo interval, conditional on not having been stunted at the previous time period (2 mo earlier), as a function of the set of independent variables described above. We used a discrete time framework because data were obtained on the children at bimonthly intervals, and it is not possible to ascertain the exact day on which a child became stunted. The analysis sample consisted of 2859 infants with length data at 2 mo of age. We began our analysis of stunting at 2 mo of age because the determinants of stunting at birth (intrauterine factors) are not comparable to those (such as feeding and morbidity) that the child encounters in the postnatal environment. However, the child's status at birth is represented in the model by birth weight.

The estimation of the model for stunting is likely to present a set of noteworthy statistical problems. Repeated length measures of each child are highly intercorrelated. Furthermore, there are underlying characteristics of the child which we cannot observe or measure, but which affect the likelihood that the child will become stunted. These characteristics, referred to as unobserved heterogeneity, reflect a child's genetic potential or resistance to disease, and represent complexities which our model had to address.

Biases are likely to occur when the unobserved factors related to stunting are also associated with the independent variables in the model. For example, the innate healthiness or frailty of the child may influence how the mother chooses to feed the child and how much care she takes to avoid exposure to pathogens. If this is true, then the feeding and morbidity variables could be correlated with the unobserved factors that also influence stunting, that is, they are endogenous to the model. Unobserved heterogeneity confounds the relationship of stunting with feeding or diarrhea, potentially leading to a biased estimate of their effects (see Briscoe et al. 1990, Cebu Study Team 1991, VanDerslice et al. 1994 for examples and further discussion of the consequences of ignoring endogeneity). It is not possible to control for the effects of this potential confounding by including the confounder in the model, because in this case, the confounder is not observed.

One approach taken by economists to correct for endogeneity involves joint estimation of the main equation of interest (stunting) and equations for potentially endogenous independent variables (in this case, birth weight, feeding, health care and morbidity variables). The equations for the endogenous variables are specified only with exogenous household, community and individual variables, and are estimated using methods appropriate to their form (logit for categorical variables and linear regression for continuous variables). Correction for unobserved heterogeneity was achieved using the Heckman and Singer (1984) approach. Full details of this method and its application to the estimation of mortality in the same sample are presented in Guilkey and Riphahn (1996).

Once the stunting model was estimated, we simulated effects of different common scenarios on the likelihood of stunting. We calculated the cumulative predicted probability of stunting for the following: 1) males vs. females, 2) low birth weight (<2500 g) vs. normal birth weight infants, and 3) selected morbidity-feeding combinations. Because inappropriate feeding practices and high morbidity tend to occur together, we developed risk profiles based on feeding and morbidity profiles characteristic of some communities of Cebu. The "high morbidity" profile is represented by the following: 1) a 50% chance of diarrhea and a 33% chance of febrile respiratory infections, levels which are about twice the mean prevalence rates in the sample; 2) inappropriate feeding, defined as not breast-feeding and consuming only 80% of the age-specific Filipino RDA for energy; and 3) no preventive health care. This set of conditions is common in urban squatter settlements of Cebu. In contrast, the "low morbidity" profile is represented by: 1) the mean prevalence of diarrhea among fully breast-fed infants for the first 4 mo of life (3% at 2 mo, and 6% at 4 mo) and a prevalence of 10% in the months thereafter, and a 7% prevalence of febrile respiratory infections, which is about half the mean prevalence in the population; 2) appropriate feeding which included full breast-feeding in the first 4 mo of life, followed by mixed feeding with energy intake from supplemental foods at levels estimated to meet the age-specific RDA; and 3) adequate levels of preventive health care. This profile, though infrequently seen in the Cebu sample, represents optimal health care and feeding practices, which in turn, would be associated with lower morbidity (Cebu Study Team 1991 and 1992).

RESULTS

Table 1. Sociodemographic characteristics of CLHNS infants who formed the analysis sample1 [View Table]

The prevalence of stunting was very high among Cebu children, with rural children worse off than their urban counterparts. By 12 mo of age, 37.7% of rural and 35.1% of urban children were stunted. By 24 mo, 68.8% of rural and 61.9% of urban children were stunted. Overall, only about one third of sample children never had a length for age Z-score less than  2 at any point during the first 2 y of life. Furthermore, the sample children were short relative to a representative national sample of over 25,000 Filipino children. For example, at 12 mo of age, about 17% of Cebu males and 10% of females were below the Filipino 5th percentile of length-for-age (Florentino et al. 1992). Stunting tends to be highly persistent. Of children stunted by 6 mo of age, 96% were also stunted at age 2. 

The age distribution of new cases of stunting is presented in Figure 1. The peak in new cases occurred at 8 mo in males and females, with a decline, among males, in the number of new cases thereafter. During the first year, the number of new cases in males consistently exceeded that among females. During the second year, however, there were more new cases of stunting among females, with a peak at 16 mo of age.

In an attempt to pinpoint when linear growth retardation occurs relative to our measurement of incident stunting, we compared the length velocities of infants who became stunted in each 2-mo interval with those of infants who did not become stunted in that interval or earlier. Results are presented in Table 2. In each case, length velocities were lower for incident cases of stunting only in the interval immediately prior to our identification of stunting. For example, compared with infants not stunted up to age 8 mo, those whose first became stunted between 6 and 8 mo had comparable length velocities from birth to 6 mo, but dramatically lower velocity from 6 to 8 mo.

Table 2. Length velocities of Cebu infants, birth to 12 mo of age: a comparison of incident cases of stunting with infants not stunted by specified ages [View Table]

Table 3. Coefficients, standard errors and t statistics associated with independent variables in the stunting hazard model1 [View Table]

The cumulative predicted probability of stunting calculated from the means of all independent variables, and under selected morbidity and feeding scenarios is graphed in Figures 2 and 3. Results are grouped into categories representing constitutional factors present at birth, morbidity, feeding and combined feeding-morbidity effects.

The effect of the sex of the child was highly significant, with opposite signs on the coefficients for y 1 and 2 of life. This is consistent with incidence data presented in Figure 1, and is also shown by the convergence in the likelihood curves estimated for males and females in Figure 2.

Preventive health care protects infants from stunting. The major components of preventive care include maternal education, and immunization against diphtheria, pertussis and tetanus (DPT) and measles.

DISCUSSION

Second, we have singled out children whose process of growth retardation leaves them "stunted" and have ignored those children who may be experiencing growth retardation, but who do not end up with length-for-age Z scores less than 2. It is highly likely that the same growth-retarding influences affect infants of all starting sizes, and thus our results should be generalizable to less severe cases as well. However, this assumption will require empirical testing.

The effects of birth weight on stunting were strongest in the first 6 mo, then diminished. This is consistent with the structure of the hazard model, and the observed persistence of stunting in the sample. Once an infant becomes stunted, he or she is censored, and makes no subsequent contribution to the estimates. LBW infants have an increased likelihood of becoming stunted early on, and are no longer in the sample "eligible" to become stunted at a later age. However, when we look at the cumulative probability of stunting, we see a strong persistent effect of LBW. This is consistent with earlier work, which presents length curves for LBW vs. normal weight infants in the first year of life, and shows significant differences in length at 12 mo among infants who were LBW vs. normal birth weight (Adair 1989). The initial low probability of stunting related to LBW (despite the high percentage of stunting among LBW infants at birth) occurs for several reasons. First, length velocities of LBW infants are significantly higher than normal weight infants (Adair 1989); thus many LBW infants stunted at birth are not stunted at 2 mo of age. Furthermore, the model examines the effects of LBW, controlling for a number of other factors also related to LBW. We recently showed significant effects of LBW on infant feeding patterns, with LBW infants less likely to initiate breast-feeding and more likely to be weaned early (Adair and Popkin 1996).

Past literature has been quite consistent in showing a negative effect of diarrhea on growth (Adair et al. 1993, Black et al. 1982 and 1984). In addition to the effect of diarrhea, we also showed a strong effect of febrile respiratory infections on growth. Febrile infections may increase energy and nutrient needs, and suppress appetite, thus reducing growth. The lack of any significant age interactions with respiratory and diarrheal infections suggests that morbidity is equally important at all ages.