There is no printed copy available to order.
Abstract:
Stunting prevalence is a core indicator of
child health that is conventionally estimated
as the proportion of children with height-
for-age z-score (HAZ) values below -2
standard deviations (SDs), based on the World
Health Organization (WHO) growth standards.
Despite its widespread use in public health,
stunting prevalence is conceptually
problematic because it is often used to
identify a subgroup of the population that is
affected by undernutrition, rather than being
correctly interpreted as a characteristic of
the entire population. From a statistical
perspective, stunting prevalence is a measure
of location (MoL) of the HAZ distribution of
an observed population relative to the
distribution of a healthy standard
population. A higher stunting prevalence
usually represents a downward whole-
population shift of the HAZ distribution,
whereby even the tallest children in the
population are shorter than expected.
Moreover, stunting prevalence is based on
cut-off values at the tails of the HAZ
distribution and may therefore be more
sensitive to imprecise data on height and
date of birth than MoLs of central tendency
(e.g., means, medians). We hypothesized that
other linear growth metrics based on child
height data may be less sensitive (i.e., more
robust) to distribution asymmetry or the
presence of influential outliers and
therefore suitable alternatives or
complementary approaches for describing the
location of an observed HAZ distribution
relative to international standards.
The objectives of this study were to (1)
identify and describe a range of candidate
linear growth metrics that could be used as
alternative or complementary indicators for
assessing population childhood linear growth
and nutritional status and (2) assess and
compare these potential metrics of child
linear growth based on the relative strengths
of their associations with other key
population indicators and on the robustness
of these associations against variations in
anthropometric data quality. Height and date
of birth data of children under 5 years of
age from Demographic and Health Surveys (DHS)
from 64 countries (2000 to 2018) were used to
generate two types of linear growth metrics:
estimates of descriptive statistics based on
observed distributions (e.g., MoLs such as
means and stunting prevalence) and regression
model-derived estimates (e.g., predicted
means at discrete ages or slopes of decline
within a defined age range). DHS data were
also used to generate indices for
anthropometric data quality based on
principal component analysis.
Correlations between each candidate linear
growth metric and stunting prevalence among
children under 5 were compared using the
absolute value of Pearson correlation
coefficients. Absolute values of Spearman
rank correlations were used to compare
pairwise associations between each linear
growth metric and each of six population
indicators of health and well-being (e.g.,
under-5 mortality [U5M], gross domestic
product). Data quality was measured using
indices composed of either three or six
individual anthropometric data quality
indicators (referred to as the 3Q index and
6Q index, respectively). Relationships
between the metrics and population indicators
were assessed using Spearman rank
correlations (for a subset of three
indicators) and linear mixed effects models
(for all six indicators) to test their
robustness against variations in
anthropometric data quality. All analyses
were performed using four approaches for
identifying (i.e., flagging) and excluding
HAZ outliers: (1) no flagging, and therefore
no exclusions; (2) less restrictive flagging,
which excluded HAZ values <-9 SDs and >+9 SDs
from the age/sex-specific WHO standard
median; (3) WHO flagging, which excluded HAZ
values <-6 SDs and >+6 SDs from the age/sex-
specific WHO standard median; and (4)
Standardized Monitoring and Assessment of
Relief and Transitions (SMART) flagging,
which excluded HAZ values <-3 SDs and >+3 SDs
from the observed sample mean.
Results showed that descriptive linear growth
metrics and model-derived predicted HAZ
values at 2 years, 3 years, and 5 years were
strongly correlated with stunting prevalence,
with absolute values of the Pearson
correlation coefficients exceeding 0.90
across all flagging approaches. Pearson
correlations for model-derived slopes of HAZ
and height-for-age difference (HAD) from
birth to 3 years of age ranged from moderate
to strong (absolute values =0.46 and =0.77,
respectively). Predicted HAZ at birth had one
of the weakest coefficients and the largest
range of correlations across flagging
approaches (-0.35 to -0.68).
Spearman rank correlation coefficients for
relationships between linear growth metrics
and population indicators (i.e., absolute
values) ranged from 0.01 to 0.73 with little
variation across flagging approaches for each
metric. Correlation strengths of the
descriptive metrics tended to cluster near
the midpoint of this range, with stunting
prevalence generally ranking higher than
other descriptive metrics. Conversely,
correlations with model-derived metrics
varied more widely; for example, predicted
HAZ at birth had the weakest correlations and
HAD slope from 0-3 years of age often had the
strongest.
In linear mixed effects models, the
associations between descriptive linear
growth metrics and any of the population
indicators were not significantly modified by
the 3Q index. When data quality was defined
using the 6Q index, significant modifying
effects were observed for associations
between stunting and the percentage of the
population with an improved source of
drinking water. The 3Q index had a
significant modifying effect on the
associations between U5M and four of the
model-derived linear growth metrics when the
WHO or SMART flagging approaches were used;
however, these interactions were not
significant when using the 6Q index. With few
exceptions, the 3Q index had no observable
modifying effect on the associations between
the model-derived linear growth metrics and
other population indicators; these findings
were similar for analyses using the 6Q index.
Correlations between linear growth metrics
and U5M were generally weaker with high
(versus low) data quality, defined using the
3Q index; results were similar when data
quality categorization was based on the 6Q
index. However, differences between high- and
low-quality strata were usually minor.
Patterns for the other population indicators
were inconsistent with the findings observed
for U5M: For gross domestic product, the
correlation strengths were greater at high
(versus low) quality for half of the
candidate metrics when using the 6Q index;
for maternal education, correlation strengths
were generally lower at high (versus low)
quality using the 3Q index (similar to U5M),
but this pattern was not consistent for many
linear growth metrics when using the 6Q
index.
In conclusion, numerous alternative linear
growth metrics may be derived from the same
population-level child height data that are
widely used to estimate stunting prevalence.
The majority of these metrics correlate
strongly with stunting prevalence and may
therefore have no substantial empirical
advantages. Findings to date suggest that
some of the alternative metrics may
outperform stunting prevalence in terms of
the strength of their correlations with other
important population indicators. However, it
was methodologically challenging to assess
the sensitivity of metric performance to
variations in anthropometric survey data
quality. Therefore, findings so far are
insufficient to lead to a recommendation to
adopt one or more of the alternative linear
growth metrics for use in tracking country or
regional improvements in child growth. In
further research, we will apply a revised
methodological approach with a more
comprehensive set of alternative metrics to
further our understanding of the potential
application of alternative linear growth
metrics for assessing and tracking child
health and nutritional status in low-and
middle-income countries.
Alternative and Complementary Metrics of Linear Growth for Tracking Global Progress in Child Nutritional Status (PDF, 4614K)