Nonlinear Effects of Altitude on Child Growth
in Peru – A Multilevel Analysis
Growth at high altitude has been the object of many investigations after experimental
studies on animals showed that hypoxia at high altitude slows growth. Many studies have
also looked at the Andean populations and found different results. Even though a few
studies find that individuals living at high altitudes are smaller than the ones living at low
altitudes, a significant group of studies does not reveal such a clear relationship. This
study focuses on Peru, a country characterized by a diverse territory, great altitude
variations and a population with a wide socioeconomic gradient. The present analysis
differs from previous studies in three ways. First, in an attempt to reconcile the main
findings of the biological literature with the economic models of child health, it explores
the relationship between altitude and child health within a multivariate framework.
Second, it benefits from a large spectrum of altitude data and does not concentrate on one
or two isolated villages. Third, it takes into account the cluster nature of the data and
controls for correlation of variables in the same cluster through multilevel statistical
modeling. After controlling for characteristics of the children, families and communities,
the data show a significant nonlinear relationship between altitude and child nutritional
status. Peruvian children living at medium/high altitudes appear to be worse off than
children living at extremely high altitudes, where the negative effect of hypoxia on
growth could be compensated by other favorable health and environmental conditions.
World Bank Policy Research Working Paper 3823, January 2006
The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the
exchange of ideas about development issues. An objective of the series is to get the findings out quickly,
even if the presentations are less than fully polished. The papers carry the names of the authors and should
be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely
those of the authors. They do not necessarily represent the view of the World Bank, its Executive Directors,
or the countries they represent. Policy Research Working Papers are available online at
Acknowledgements: we are grateful for helpful suggestions and comments on an earlier
version of this paper from Chris Barrett and Harold Alderman.
1. INTRODUCTION: MOTIVATION AND OBJECTIVES OF THE PAPER
1.1. Altitude and growth
Growth at high altitude has been the object of many investigations after experimental
studies on animals showed that hypoxia at high altitude slows growth (Gordon et al.,
1943 and Moore and Price, 1948).
Previous studies looked at the effects of altitude on human growth exclusively from
biological or physical anthropological perspectives. Studies on human growth proved that
altitude influences anthropometric outcomes but did not show uniform and
uncontroversial results on the influence of altitude on height. Pawson (1977) does not
find a significant difference between the high altitude Sherpa and the low altitude Tibetan
(living in Nepal). Malik and Singh (1978) find that in India children living at high
altitude are taller than low altitude children in late adolescence but the opposite is true for
children in early adolescence. Clegg et al (1972) show that in Ethiopia children living at
high altitude are taller than children living at low altitude, maybe because of the higher
prevalence of infectious diseases (such as malaria) at low altitude. Many studies have
also looked at the Andean populations and found different results. Even though a few
studies showed that individuals living at high altitudes were smaller than the ones living
at low altitude (Haas, 1976; Frisancho and Baker, 1970; Beall et al., 1977, Mueller et al.,
1978), a significant group of studies could not reveal such a clear relationship (Hoff,
1974; Pawson, 1977, Clegg et al. 1972, Frisancho et al, 1975). In general, consensus
seems to exist only with regard to the increased chest size of high altitude populations.
There are many ways through which altitude can have an effect on child health. Higher
altitude is often associated with more difficult transportation, which, on one side, can lead
to higher food prices and have an impact on the diet of the children and, on the other side,
to more difficult access to health facilities. Moreover higher altitude can be associated
with worse crop outcomes and have an impact on the diet of the population and on their
available resources. Unfortunately we were unable to find data on prices, transportation
or crops for our study. We are therefore unable to disentangle the different mechanisms
through which altitude can affect child health but notwithstanding we can say something
about their relationship at the national level. Understanding the role of altitude can shed
some light on the effect of other factors on anthropometric outcomes. Indigenous
population is often found at disadvantage even after controlling for income and
infrastructure. That could be due to unobservable factors such as social exclusion or to
the correlation of ethnicity and altitude (Alderman et al, 2000). Being able to control for
altitude allows disentangling the two separate effects.
The majority of studies looking at the health effects of altitude belong to the biological
literature and, as a typical research strategy, used to contrast native populations at
different altitudes (Baker, 1978). Those studies were therefore able to take into account
ethnic background (and genetic variations for adolescents and adults) but often looked at
the growth profile of a person as result of a single environmental factor in isolation rather
then viewing it as the result of different factors and their interaction. In particular, those
studies could not control for other environmental factors such as disease prevalence,
nutritional intake and maternal care that have been proved to contribute to large
differentials in child health (Mosley and Chen, 1984). Many of these factors would be
endogenous in a classical study of determinants of health but could be accounted for by
controlling for household income, household resources, parents’ education and
availability of infrastructure.
The study focuses on Peru, a country characterized by a very diverse territory and great
altitude variations. The central portion of Peru includes the great mountain and plateau
region of the Andes, with numerous peaks rising to over 6,000 meters and with extensive
plateau districts between 3,000 meters and 4,300 meters. There is a very narrow coastal
plain on the Pacific shore, while to the east of the Andes, the land drops steeply to the
forested lowlands of the Amazon basin.
In what follows we look at the relationship between child growth and altitude within a
classical Beckerian model of the family. The present analysis differs from previous
studies in three ways. First, in an attempt to reconcile the main findings of the biological
literature with the economic models of child health, we explore the relationship between
altitude and child health within a multivariate framework. Second, we benefit from a
large spectrum of altitude data and do not concentrate on one or two isolated villages.
The majority of altitude studies on Peru, for example, used a sample of individuals from
the rural highland community of Nuñoa (4000 mt). Later studies (Leonard et al. 1990)
showed that Nuñoan are among the smallest of all Andean population and it would
therefore be misleading to use them as evidence of the negative effect of altitude. It has
been shown that a combination of both ecological and sociological constraints on food
availability puts people in Nuñoa under nutritional stress as much as it does for other
population in developing countries around the world. Finally, and more importantly, our
analysis takes into account the cluster nature of the data and control for correlation of
variables in the same cluster through a multilevel analysis.
1.2. Hierarchically Clustered Data
The data we are using for the analysis of child health in Peru present a hierarchical
structure: factors affecting health outcomes arise from different levels of aggregation: the
outcome of interest, child health, takes place at the individual level and is influenced by
higher level characteristics which do not vary between individuals of the same group
(household and community). In general household level or community level factors are of
great interest because they can often be influenced by policies to affect individual level
variables. In previous studies, higher level determinants of child health have always been
observed and included in the analysis. The inclusion of multilevel factors among the
determinants does not change the interpretation of the effects of those factors. Problems
may arise if higher levels unobserved characteristics influence the lower level variables.
The OLS estimator is as efficient as the maximum likelihood estimator only when the
community level covariates and the household level covariates are uncorrelated with all
the individual covariates (Angeles, Guilkey and Mroz, 2002). In general, however, the
OLS estimator understates the true standard errors.
Researchers have often adopted fixed effects models to estimate nutrition models and
control for unobservable variables at the cluster level. The main difficulty in using fixed-
effect models is that if the fixed effect is differenced away, then the effect of those
variables that do not vary within a cluster will be lost in the estimation process. And that
is particularly problematic in our analysis of the relationship between altitude and growth
since altitude data are available only at the cluster level. We use a multilevel analysis
model because of the clustered nature of the data and because we want to incorporate
context in our analysis in order to study the impact of altitude.
2. MODEL AND EMPIRICAL STRATEGY
2.1. Beckerian model of the household
The analysis that follows is based on a standard Beckerian model of the household.
Households are assumed to maximize a utility function defined over consumption of a
composite good (the vector of consumption goods of different individuals in the
household), household members’ leisure and their health. Households maximize their
utility function under several constraints, including a time constraint, a budget constraint,
and a biological health production function. The health production function relates the
health status of the child to his or her past health conditions and a set of inputs chosen by
the household (including food intake, breastfeeding, utilization of health facilities, and
the time dedicated by the mother to health related activities), a set of exogenous
characteristics (such as the child’s age and gender), a set of household characteristics
including parents’ health and education, their investment in child care, household
resources and the available facilities.
The family optimization problem can be solved to yield a reduced-form equation for
health outcomes in which child health depends only on exogenous individual, household,
and community characteristics:
Ni = n(Ci, Ch, Cc, εi),
where: Ni is the height-for-age z-score for child i; Ci are the individual characteristics of
the child, including age and sex; Ch are household characteristics that incorporate
measures of family background, including resource availability, parents' health, and
parents' skills, measured generally by their level of education, and whether the father is
absent from the household; Cc are community characteristics, including altitude level, the
availability of health services, the state of infrastructure such as water and sewage, and
other community characteristics that affect child health through the proximate
determinants; and εi is an individual specific random disturbance associated with the
anthropometric outcome of the child and assumed to be uncorrelated with the C variables.
Estimation of the reduced-form anthropometric function does not provide information on
the biological mechanisms responsible for children's growth deficits, but it does provide a
consistent statistical framework within which to estimate the impact on children’s health
and nutrition of individual, household and community exogenous variables that are
generally open to policy intervention. The parameter estimates of the coefficients in the
reduced-form equation can be interpreted as the full effects of exogenous covariates, that
is their effects not mediated by the proximate determinants.
2.2. Multilevel Analysis
Traditionally reduced form models are estimated with ordinary least square techniques.
One of the critical assumptions of OLS models is the independence of disturbance. But
in cluster samples, such as our data, observations are not independent: the growth
experience of children within the same community may be similar, especially if they
come from the same family. OLS estimates of this type of data can therefore result in
inefficient estimates of the parameters and underestimated standard errors. By ignoring
the hierarchical structure of the data we are ignoring a significant and interesting
community effect. As evident in the graph below, that represents the average z-score by
community, the average z-score varies substantially across communities (figure 1).
Figure 1 - Cluster mean of HAZ against cluster
Source: Author’s calculation using Peru DHS 2000 data.
Multilevel models are random effect models, which take into account the hierarchical
nature of the data. Individuals but also households and communities are the unit of
analysis. (Kreft, de Leew, 1994). In these models the greater homogeneity of
observations in the same group is modeled by adding a random effect at each cluster:
zijk = β’xijk + δk + μjk + εijk ,
where zijk is the height-for-age z-score for the ith child of the jth family in the kth
community; β is a vector of regression coefficients corresponding to the effects of fixed
covariates xijk, which represent observed characteristics of the child, the family and the
community; δk is a random community effect that represents the deviation of community
k’s mean z-score from the grand mean; μjk is a random family effect that represents the
deviation of family jk’s mean z-score from the mean of community k; and εijk is an
individual error term that represents the deviation of child ijk’s z-score from the mean of
The random effects μjk and δk represent unobserved family and community factors shared
between siblings and between children living in the same community respectively.
Anthropometric outcomes of children living in the same community (but not in the same
family) are correlated because they share the random effect δk, and anthropometric
outcomes in the same family are correlated because they share the random effects δk and
The error terms δk, μjk, and εijk are assumed to be normally distributed with mean zero and
variances σ2c, σ2f, and σ2i respectively. If σ2c is zero, observations in the same
community (but not in the same family) are independent. If σ2f is also zero, observations
belonging to the same family are also independent. If σ2c and σ2f are not zero, the
observations are correlated and the OLS assumption of independence does not hold. The
variances of the random terms are the additional parameters estimated by variance-
components models as compared to OLS linear regression models. To the extent that the
greater homogeneity of within-cluster observations is not explained by the observed
covariates, σ2c and σ2f will be larger. To evaluate the appropriateness of our model we
therefore test whether the variances of the random part are different from zero over
families and over communities.
In these models the coefficients can be fixed (variance component models) or random
(random coefficient models) and the choice can be made separately for every coefficient.
For example, we can think of the effect of household resources on child health as varying
from community to community instead of being fixed across communities. This would be
equivalent to assuming that the slope of asset index is also a random effect βi, which we
assume normally distributed with mean β and variance σ2β, where β is now the average
effect of household resources across the surveyed communities.
As mentioned before, multilevel models don’t assume that observations of different
individuals in the same family or in the same community are independent. As a matter of
fact, from the resulting estimates we are able to assess the extent to which child health is
correlated within families and within communities, before and after we have taken into
account the effect of the observed covariates xijk. The Intra Class Correlation (ICC)
coefficient, in particular, is used to assess the amount of covariation between
observations belonging to the same group. Zero correlation means that the observations
are independent.1 When the correlation is different from zero, it is more appropriate to
1 Because we are assuming that the errors are normally distributed.
use random effect models. Note that even small values of ICC have been showed to lead
to type I errors that are much larger than the normal alpha level of 0.05 (Hox and Kreft,
The ICC coefficient describes the proportion of variation that is attributable to the higher
level source of variation. The correlations between the anthropometric outcomes of
children in the same community and in the same family are respectively2:
ρc = σ2c /(σ2c + σ2f + σ2i)
ρf = (σ2c + σ2f )/(σ2c + σ2f + σ2i).
The total variability in individual anthropometric scores can be partitioned into its three
components, that is variance among: children within families, families within
communities, and communities. Multilevel models allow us to evaluate whether child
health can be attributed to individual differences, differences between households or
structural differences between communities. Finally, by including covariates measured at
the individual-, household-, and community-level, variance-components models enable
us to explore the extent to which community differences in average height-for-age z-
scores are accountable for by factors operating at each level.
The Peru Demographic and Health Survey of 2000 (DHS 2000) is a nationally
representative, probabilistic, self-weighted, stratified survey that covers 28,900
households and 27,843 women between 15 and 49 years. Two questionnaires were
covered in the survey. The household questionnaire collected information on the
characteristics of the households like economic activity, assets and infrastructure and
achieved education level. The individual questionnaire collected a series of information
on women’s background, their reproductive and fertility history, and their health,
including anthropometric measurements of their children. Our investigation was
restricted to 11,585 children between 0 and 60 month whose anthropometric
measurements were available.
The data used in the study are characterized by a hierarchical structure: individuals are
nested within households3 and households are nested within communities (clusters). The
2 The intra family correlation coefficient takes into account the fact that children in the same families also
live in the same communities.
3 The DHS survey collected information for every mother in the family. In theory, we could have therefore
set up a four level model with random effects at the individual, mother, household and community level.
But the limited number of families with more than one mother (3%) and the reduced size of the household
level cluster did not make it possible. For those families with more than one mother we have collapsed
mothers’ information at the family level. To do so we kept the age of the oldest mother and the education
level of the most educated mother, based on the assumption that experience and education of the woman in
the family can have a positive externality effect on all the children in the family. We defined as non-
indigenous those families where there was at least one non-indigenous mother, again based on the
11,585 children of whom we have observations come from 8,925 families distributed
around 1,325 clusters. Sample sizes averaged about 8 children per community, and 1
child per family. The fact that the clusters are very small (on average 6 families per
community) justifies even more the adoption of multilevel models. Child health is
measured by the height for-age z-score, an indicator of stunting. Stunting represents the
accumulated consequences of retarded skeletal growth. It reflects the cumulative effects
of the many different insults that children in developing countries experience in the
uterine and preschool years and is frequently found to be associated with poor overall
economic conditions. Moreover stunting is largely irreversible after 2 years of age and for
this reason is considered an accurate indicator of long-term chronic malnutrition in early
childhood (Keller, 1983; Behrman and Deolalikar, 1988; Strauss and Thomas, 1998,
Martorell and Scrimshaw, 1995). The average height-for-age z-score of children in the
sample is –1.17 for boys and –1.18 for girls. 25.2% of the boys and 25.6% of the girls in
the sample are stunted.
4. ALTITUDE AND MALNUTRITION: DEBUNKING AN OLD MYTH?
Most of the empirical studies that analyzed the relationship between altitude and human
growth compared anthropometric data of genetically similar populations living in two
different villages or communities, one at high altitude (above 3500 or 4000 meters) and
the other at low altitude (at sea level or below 1500 meters). The results of the different
studies have been equivocal with respect to specific effects of hypoxia on growth. As
suggested by Mueller et al. (1978), some of the lack of consistencies in the results of
human studies could reflect the failure to keep constant other factors that vary with
altitude, beside hypoxia. The purpose of the following analysis is to explore the
relationship between growth and altitude from a different perspective, using a spectrum
of altitude data that covers the whole country and estimating the relationship within a
household demand framework that allows taking into account (directly and indirectly) the
most important determinants of child health.
assumption of positive externalities coming from a Spanish-speaking person in the family (ethnicity is a
Figure 2 - Average Height-for-Age Z-Scores by Altitude Level (1994, 1996, 1997,
Average haz, 1994
Average haz, 1997
Altitude level (mt)
Average haz, 1996
Average haz, 2000
0 1000 200030004000
Source: Author’s calculations using data from ENNIV94, ENNIV97, DHS96, DHS00.
Figure 2 shows the average height-for-age z-score per altitude level in four different
periods in time. The first striking observation is the great uniformity of the different trend
lines: despite the differences in survey methodologies, samples and sample sizes, the
relationships between altitude and malnutrition look very similar over time. One possible
explanation is that factors affecting growth at different altitude levels were unchanged
over the period 1994-2000.
The data also show the existence of a downward trend in height for age z-scores up to
3500 meters. According to the data, it seems that the smallest children live between 2500
and 3500 meters. Another notable and unusual outcome is the slight upward trend in
children’s growth above 3500 meters confirmed by all the surveys but DHS 1996.
Note that if we sampled two separate groups of children, one from communities below
1500 meters, and the other from communities above 3500 meters, as biological studies
did in the past, we would have quite likely observed that children living at higher
altitudes were smaller than children living at lower altitudes. On the other side, when
analyzing the variation in altitude and growth at the country level we find a more
complex and non-linear relationship, as illustrated by table 1 below.
In an unconditional model, where height for age z-score is regressed simply on altitude
levels without taking into account differences in other household or community
characteristics, child growth worsens at different rates with increasing altitude levels. If
we control for household resources then children living between 2500 and 3500 meters
have worse health outcomes than children living above 3500 meters and this relationship
is even stronger if we also control for family’s background characteristics (as represented
by mother’s height).
Table 1 - Non-Linear Relationship Between Child Growth and Altitude
altitude >= 1500, <2500 mt -0.46529
altitude >= 2500, <3500 mt -0.79977
altitude >= 3500 mt -0.83976
Household asset index
Mother's height (cm)
Absolute value of t-statistics in parentheses
* significant at 10%; ** significant at 5%; *** significant at 1%
Figure 3 and 4 below show growth patterns by age groups for four different altitude
levels. Nutritional status of children living at higher altitude appears to be worse that that
of children leaving at lower altitudes, but once again children living above 2500 meters
are not necessarily worse off than children living between 1500 and 2500 meters.
asi & mheight
Figure 3 -Female Stunting by altitude and age group
0-67-12 13-1819-24 25-3637-48 49-60
Proportion of Stunted
Source: Author’s calculation using Peru DHS 2000 data.
Figure 4 -Male Stunting by altitude and age group
0-67-12 13-1819-24 25-3637-48 49-60
Proportion of Stunted
Source: Author’s calculation using Peru DHS 2000 data.
This complex relationship between altitude and child growth could be the outcome of a
combination of opposite forces operating at the same time. The negative effect of hypoxia
on growth at high altitudes could for example be compensated by the existence of more
difficult health conditions (such as higher prevalence of infections) or by worse
environmental conditions (such as pollution in urban areas) at lower altitudes. Finally, an
alternative explanation may come from variations in agricultural production at different
altitude levels and specifically by the fact that above 3,500 meters Peruvian peasants tend
to stop cultivating maize and secure land for pastoralism (Cotlear, 1986) with
consequences on their diets that is likely to be much richer in protein.
In this section we analyze the relationship between child malnutrition and altitude in Peru
controlling for the effect of different exogenous characteristics. We adopt a multilevel
model to control for clustering and to be able to add context (altitude) in the analysis.4
Malnutrition will be explained by three groups of variables: individual, household and
community -level variables. Means and standard deviations of these variables are
presented in table 2.
We have fitted several models in order to better understand the relationship between
altitude and nutritional status. We start by fitting a reference model (model 1) that
includes only age and gender of the child. Model two includes dummies for altitude
ranges to represent the non-linear relationship between altitude and child growth. Model
three is a full model that includes all the exogenous covariates and in model four we add
the interaction between altitude and migration history of the mother.
4 All the computations have been carried out using the MlnWin computer package (Rasbash et al., 2000).
Table 2 - Means and Standard Deviations
height for age z-score
age in months
age of the mother - years
mother has completed primary education
mother has completed secondary education
mother has completed post-secondary education
mother is indigenous
father has completed primary education
father has completed secondary education
father has completed post-secondary education
Household size - number of people
number of children age 0-5
number of women age 16-25
number of women age 26-65
altitude - meters
alt > 3500
urban area dummy
Proportion of hh with low qual floor
Proportion of hh with piped water
Proportion of hh with water from river/stream
Proportion of hh with flush toilet
Proportion of hh with no toilet
Proportion of hh with electricity
Proportion of hh with tv
Proportion of hh with fridge
Proportion of hh with phone
Migrated before child'd birth - dummy
Source: Author’s calculation using Peru DHS 2000 data.
The parameter estimates are presented in table 3.
Estimates of the reference model presented in column 1 are a useful preliminary step as
they provide information about the outcome variability at each level.5 A substantial part
of the total variance (42%) is attributable to family and community level variation in the
height-for-age z scores. The individual variation is almost three-times higher than the
variation at the community or family level. Part of this variation is due to measurement
5 The estimation method is Iterative Generalized Least Squares (IGLS) and convergence is judged to have
occurred when all parameters between two iteration have changed by less than a tolerance of 10-2.
errors originating by either misreported height or age of the child. Interestingly, there is
more variation between communities than between families, possibly reflecting the great
geographical differences of the country. As a matter of fact, differences in altitude levels
account for an important proportion of the variation between different communities: if we
add altitude to the reference model (model 2) we immediately observe a significant
reduction in the community variability (by 31%).6 In model 2 the effect of altitude on
child growth is negative and significant. Without controlling for further exogenous
characteristics, the model shows that child’s nutritional outcome worsens monotonically
as altitude increases.
As a next step we control for the relevant factors at the individual, household and
community level (model 3, 4 and 5). Note that the individual random effect remains very
high even after including other covariates as the only individual level variables used in
the analysis are age and sex. On the other side, estimates of family variation indicate that
a good part of it (approximately 30%) is explained by the observed exogenous factors in
the model while 86% of community variation is explained.7
The effect of altitude is highly significant in all model specifications. Note that after
controlling for the individual, household and community characteristics, the relationship
between altitude and child nutritional status becomes non-linear. Peruvian children living
at medium/high altitudes (between 2500 and 3500 meters) appear to be worse off than
children living at extremely high altitudes (above 3500 meters). These results should be
interpreted with some caution as the altitude variables could be capturing other
geographical and environmental characteristics not included in our model. Unfortunately,
we were unable to find data on prices, soil characteristics, transportation: variables that
are often correlated with altitude and may have helped disentangling the different
mechanisms through which altitude affects child health. The large proportion of
community level variance explained by the model (85%) suggests on the other side that
most of the variables explaining differences between various communities have been
In model 4 we control for the migration history of the family and test whether the fact
that the mother had been living at high altitudes before the child’s birth had an impact on
the nutritional status of the child. We expect that, if there is an hypoxia effect, longer
exposure to high altitudes would have a negative effect on nutritional outcomes. To do
6 We formally test that adding altitude dummies improves the fit of the model by using a likelihood ratio
test. The difference between –2LL for the reference model and the model with altitude dummies can be
compared to the χ2 distribution on 3 degrees of freedom (we are including three extra parameters). The
difference is 307. This value is highly significant (P-value: 0.000)
7 Note that in a multilevel model such as the one above, the inclusion of an extra explanatory variable with
a fixed coefficient will generally change either (or all) the level 1, level 2 and level 3 variances. If, for
example, the variable is measured at level 2 then the general effect will be to reduce the level 2 variance but
leave the level 1 variance unaffected. When the variable is uncorrelated with the level 2 residuals we would
not expect any reduction in the level 2 variance. If instead the variable is measured at level 1 then the level
1 variance will generally be reduced. If is a cross-level interaction term then both variances can increase
(because it practically becomes a level 1 variable). If x is measured at level 1, however, then the level 2
variance can increase since the level 2 residuals are now conditioned on a further variable.
so, we interact altitude dummies with a dummy that is equal to one if the mother moved
to the actual location before the birth of the child. We find that the only significant
interaction is that between the migration dummy and the dummy for living between 1500
and 2500 meters. It would seem that altitude affects nutritional status of the child
independently of the duration of the exposure, perhaps because of other effects that
altitude may capture. The migration dummy taken by itself (“mother moved to the village
before the child’s birth”) is on the other side significant and positive in all models,
indicating that children whose mothers have been living in the current village before
giving birth are better off than children whose parents moved only recently and possibly
that hypoxia is not the driving mechanism. The result suggests that the process of
adjustment and adaptation to a new community and its social life can have a negative
effect on the nutritional outcome of the child.
In model 5 we control for the health characteristics of the regions that children live in by
including the regional infant mortality rate. We do that for two main reasons. First, we
are trying to indirectly capture the health characteristics of the area, including availability
and quality of health services. Second, we are worried about a potential selection bias:
that the children we observe at high altitudes are those who survive the difficult health
and environmental conditions and therefore the healthiest ones. Controlling for mortality
slightly affects the estimates of the altitude coefficients but does not change the main
results. Living at medium-low altitudes (1500-2500 meters) is no longer significantly
different from living at very low altitudes. Moreover, as consistent with previous
estimates, children living at very high altitudes are better off than children living at
The other results of the model are consistent with the standard literature of economic
determinants of malnutrition. Unsurprisingly for Latin America, there is no evidence of
gender differences in child growth. Also, results confirm the cumulative nature of
stunting, which increases monotonically especially during the first two years of the child
(weaning age) and then tends to stabilize.
A mother’s height has a highly significant impact on the nutritional status of the child.
The effect has been interpreted in the past as representing unobserved family background
characteristics in addition to capturing genetic influences and the mother’s health
endowment (Horton, 1986; Barrera, 1990; Thomas, Strauss, and Henriques, 1990).
Moreover, the mother is recognized to have an environmental effect on child nutritional
status through the womb.
In all of the models, mother’s education has a positive and significant effect on children’s
nutritional status. And it is only secondary and post secondary education of the mother
that matter. Education often helps mothers to understand how to deal with nutrition,
disease, and sanitation most effectively. In addition, education influences other
socioeconomic characteristics like the age at which women marry, the number of children
8 In estimating the response of child nutritional status here, the mortality rate faced by each family is treated
as predetermined as we are using average regional rates. In doing so we undoubtedly neglect some
feedback effect from malnutrition to mortality.
they have, and their status within the community. The fact that a father has completed
primary education, on the other side, seems to be related with worse nutritional status of
the children. The same unexpected finding has been reported in previous studies of
determinants of child nutritional outcome in developing countries (Skoufias, 1998).
We observe no significant effect of a mother’s age, in line with previous analysis of child
malnutrition in Peru (Ruggeri Laderchi, 2001).
Indigenous children are significantly more malnourished than non-indigenous children.
There are a number of characteristics that may contribute to lower indigenous children’s
height relative to non-indigenous children, such as the fact that they tend to live in low-
income households, in rural areas and have less-educated parents. After controlling for
income and other household and individual characteristics, ethnicity is still an important
determinant of child growth attainment; but note that the effect of ethnicity diminishes
significantly (-32%) once we take into account for differences in altitude level9
suggesting that ethnicity captures many unobserved geographical and community
Household resource availability has a substantial and significant effect on children
nutritional status in Peru. Other household demographic characteristics, like household
size, number of children and number of adult women, also appear to be important
correlates of child growth. Children living in larger household and household with other
preschool children appear to be at disadvantage in terms of growth, suggesting the
existence of competition for resources and care.10 On the other side, the number of adult
women has a positive effect on child nutrition, especially if women are older (the effect
of the number of women between 26 and 65 years is more than double than that of
women between 16 and 25) suggesting that experience has a relevant effect on nutritional
Living in rural areas was not related to stunting once other variables were taken into
We controlled for two groups of community covariates: the environment children live in
and the degree of modernization their families are exposed to. After controlling for
altitude and for clustering in the data the only environmental characteristic that seems to
matter is the proportion of households in the community with sanitation, which has a
substantial positive effect on child’s growth. Improved sanitation is expected to be
associated with reduced exposure to infectious agents and therefore better health status.
Turning to our measures of community modernization, only the effect of proportion of
households with a television is positive and significant and robust to different model
9 The coefficient of ethnicity declines from –0.238 to –0.161 after controlling for altitude.
10 Note that the inclusion of fertility variables in the model presents an additional problem; fertility, like
child health, is part of the household decision making process. Estimates of the effect of fertility on child
health are therefore likely to suffer from endogeneity bias. Omitting the fertility variables, on the other side,
would introduce bias in the other coefficients. For this reason we decided to include the variables in the
model but to be careful in interpreting their effect on child health.