Next Article in Journal
Vitamin D, the Sunshine Molecule That Makes Us Strong: What Does Its Current Global Deficiency Imply?
Previous Article in Journal
Exploring Vegetarian and Omnivorous Approaches to Cardiovascular Risk and Body Composition
Previous Article in Special Issue
The Importance of Philanthropy Foundation for the Future Sustainability of Agriculture and Nutrition: An Opinion Study on Practical Applications, Policies, and Strategies
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Nutrients
Volume 16
Issue 13
10.3390/nu16132014
nutrients-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles thumb_up
...
Endorse textsms
...
Comment
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

The Socio-Environmental Determinants of Childhood Malnutrition: A Spatial and Hierarchical Analysis

by
Austin Sandler
1 and
Laixiang Sun
2,3,*
1
Department of Agricultural and Resource Economics, University of Connecticut, 1376 Storrs Rd., Unit 4021W.B. Young 302, Storrs, CT 06269, USA
2
Department of Geographical Sciences, University of Maryland, 2181 LeFrak Hall, 7251 Preinkert Dr., College Park, MD 20740, USA
3
School of Finance & Management, SOAS University of London, London WC1H 0XG, UK
*
Author to whom correspondence should be addressed.
Nutrients 2024, 16(13), 2014; https://doi.org/10.3390/nu16132014
Submission received: 21 May 2024 / Revised: 11 June 2024 / Accepted: 19 June 2024 / Published: 25 June 2024
(This article belongs to the Special Issue Sustainable Food Systems, Optimal Nutrition, and Health Promotion: Challenges in the 21st Century)
Browse Figures
Review Reports Versions Notes

Abstract

:
Despite a remarkable reduction in global poverty and famines, substantial childhood malnutrition continues to persist. In 2017, over 50 million and 150 million young children suffered from acute malnutrition (wasting) and chronic malnutrition (stunting), respectively. Yet, the measurable impact of determinants is obscure. We evaluate proposed socio-environmental related determinants of stunting and wasting across Kenya and Nigeria and quantify their effectiveness. We combine health and demographic data from Kenya and Nigeria Demographic Health Surveys (2003, 2008–2009, 2013, 2014) with spatially explicit precipitation, temperature, and vegetation data. Geospatial and disaggregated data help to understand better who is at risk and where to target mitigation efforts. We evaluate the responsiveness of malnutrition indicators using a four-level random intercept hierarchical generalized logit model. We find that spatial and hierarchical relationships explain 28% to 36% of malnutrition outcome variation. Temporal variation in precipitation, temperature, and vegetation corresponds with more than a 50% change in malnutrition rates. Wasting is most impacted by mother’s education, family wealth, clinical delivery, and vaccinations. Stunting is most impacted by family wealth, mother’s education, clinical delivery, vaccinations, and children asymptomatic of fever, cough, or diarrhea. Remotely monitored climatic variables are powerful determinants, however, their effects are inconsistent across different indicators and locations.

1. Introduction

Childhood malnutrition is a pernicious public health issue. It is a detrimental and significant plight responsible for 45% of all deaths among children worldwide [1]. Malnutrition not only increases child morbidity and mortality, it also inhibits cognitive, social, and financial potential [2,3]. Ongoing progress to reduce malnutrition has so far been insufficient to attain the World Health Assembly targets for 2025 and the Sustainable Development Goals for 2030 (i.e., a 40% reduction in stunting prevalence and reduce wasting prevalence to less than 5% by 2025, and by 2030 end all forms of malnutrition) [4,5]. Despite downward global trends, only 26 of 202 countries are on track to meet the undernutrition target [6,7].
Causes of child malnutrition are broadly divided into two etiological categories: illness-related or non-illness-related [8]. The focus of this paper is to evaluate the latent determinants that impact the severity and variability of non-illness-related childhood malnutrition. Non-illness-related malnutrition stems from economic, social, environmental, political, or cultural factors that decrease nutrient intake and negatively affect growth and development [9]. Malnutrition severity is measured by deterioration in various key anthropometric indicators.
Two of the most widely studied indicators are wasting and stunting. Wasting indicates a deficit in tissue and fat mass, either from weight loss or inability to gain weight. A child, aged 0 to 59 months, is defined as wasted if their weight-for-height is below negative two standard deviations from the median of the WHO Child Growth Standards [10]. Stunting indicates impeded skeletal growth. It is a measure of linear growth, representing chronic malnutrition accumulated over time. A child, aged 0 to 59 months, is defined as stunted if their height-for-age is below negative two standard deviations from the median of the World Health Organization Child Growth Standards [10].
Since the introduction of the 1990 UNICEF conceptual framework, it has been the standard for modeling the broad causes of child malnutrition. Driven in part by demand from governmental and non-profit aid agencies to successfully carry out their missions, there has been an upsurge in studies attempting to corroborate the UNICEF framework with empirical evidence and adapt it to fit new paradigms [4,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31]. Empirical evidence for non-illness-related childhood malnutrition determinants and policy interventions spans micronutrient supplementation [32], various sociodemographics [33,34,35], and climactic factors [36]. Other conceptualizations focus on different factors such as food security [37,38,39,40,41,42], risk factors [1,8,32,43,44,45,46,47,48,49], national economic growth [50,51], spatial composition [52,53,54], and utility maximization [55,56], all disseminating from a wider historical epistemology.
The UNICEF conceptual framework models child malnutrition as a hierarchical system to motivate the drivers of malnutrition [57]. The hierarchical strata include immediate, underlying, and basic classifications. Some specifications equate these strata to individual, household, and societal levels, whereby factors at one level influence other levels [29], while other specifications focus on distinguishing between proximal and distal determinants [58,59].
For example, in their report on the aggregate cross-county determinants of malnutrition, Smith and Haddad [2] identify specific subcategories of the UNICEF framework. They specify dietary intake and health status as the immediate determinants, which are influenced by the underlying determinants of food security (per capita national food availability), care for mothers and children (women’s education and women’s status relative to men’s), and health environment quality (safe water access), which are in turn influenced by the basic determinants of economic resource availabilities (per capita national income) and the political environment (democracy score).
Despite long-observed environmental effects [60] and widely anticipated links between climate change and child malnutrition, evidence regarding the nature, severity, and variability of this relationship is just beginning to emerge across spatial and temporal scales [61]. Others have found that much of the evidence for the impact of climate on childhood malnutrition is based on a few heterogeneous studies with flawed methodologies [36]. Indeed, there exists abundant opportunities in the literature for studies with wider geographic coverage and greater attention to scale. Such studies should include multiple dimensions of nutrition outcomes and quantify the spatial, social, political, climatic, and economic determinants of malnutrition.
Globally Nigeria has the second highest number of stunted children behind India. There are over 4.8 million wasted children and over 10 million stunted children in Nigeria. And Kenya provides a measure of external validity and contrast to this analysis, with over 278 thousand wasted children and over 1.8 million stunted children [10]. Including both countries adds variability in terms of malnutrition prevalence rates, governance, climate, population, economy, and culture. Kenya and Nigeria both have available health survey data with extensive temporal and spatial coverage that overlaps with the available geolocated historical climate data. Despite similar efforts in the literature in both Kenya [39,62,63,64] and Nigeria [11,12,65,66,67,68], this paper aims to be the first study to identify and quantify the non-illness-related and climatic determinants of stunting and wasting across Kenya and Nigeria through a spatially explicit hierarchical modeling approach consistent with the UNICEF [29,69] conceptual framework.

2. Materials and Methods

2.1. Health Survey Study Design

The Demographic and Health Surveys (DHS) remain the most ubiquitous resource of its kind, with more than 350 surveys in over 90 countries across 30 years [70]. To supply the primary data on child health and household characteristics, we employ the Demographic and Health Surveys (DHS) Kids Recode files and the Geographic Data files for Kenya and Nigeria of DHS-IV (1997 to 2003), DHS-V (2003 to 2008), DHS-VI (2008 to 2013). The sample includes 48,086 Nigerian children and 28,421 Kenyan children.
The sampling procedure is a multistage probability sample design, drawn from the sampling frame of the most recent census. Regions, zones, or provinces stratify national populations, and states or counties stratify regions. The final stratum contains a subpopulation from which clusters are randomly sampled. The extent of clusters vary; they can be a city block or apartment building in urban areas whilst being a village or group of villages in rural areas. Generally, geographic regions and urban or rural areas within each region partition the stratified samples of the Demographic and Health Surveys [71].
For example, the sampling frame for Nigeria partitions the country into 6 geographical regions, 36 states and the Federal Capital Territory, 774 local government areas, 8812 wards, and 665,000 census enumeration areas, each containing 48 households on average. The sample design for Nigeria DHS-VI selected 893 wards with a selection probability proportional to its population and stratified across urban and rural local government areas in each state. The sample design then selects 904 census enumeration areas from within the 893 wards and if a selected enumeration area contains less than 80 households, a neighboring enumeration area is added to form the primary sampling units or clusters. Finally, the sample design selects a fixed number of 45 households from each cluster to determine who to interview.
The probability of selecting a household is the probability of selecting the cluster multiplied by the probability of selecting the household within the cluster. Kenya DHS clusters all have 25 households whereas Nigeria DHS-VI clusters have 45 households; Nigeria DHS-V clusters have 41 households; and Nigeria DHS-IV clusters have 22 households on average. The number of clusters do not remain constant across time and space. Kenya DHS-V and DHS-IV have 400 clusters, whereas Kenya DHS-VI has 1612 clusters and Nigeria DHS-IV has 365 clusters; Nigeria DHS-V has 888 clusters; and Nigeria DHS-VI has 904 clusters. The stratified samples produce homogeneity within groups and heterogeneity between groups.

2.2. Data Measures

We employ data from three of four main questionnaires of the Demographic and Health Surveys. The Household Questionnaire characterizes the household in terms of physical amenities and a roster of the members of the household. The Biomarker Questionnaire characterizes the anthropometric measurements and biochemical indicators of eligible members of the household. The Woman’s Questionnaire contains a birth history roster of detailed health and nutrition statistics for select eligible children, in addition to characteristics about the woman. The birth history forms the basis for the Kids Recode file, a standardized module of information related to the child’s pregnancy and postnatal care and immunization, health and nutrition data [72]. The recode file is a standardized file that facilitates cross-country analysis.
We calculate anthropometric measurements, including weight (tenths of a kilogram) and height (tenths of a centimeter) from the Kids Recode files. We construct the z-sores of anthropometric indices using Stata Statistical Software 15.1 [73,74]. We input the weight and height measurements along with the sex and child’s age to calculate z-scores following the 2006 World Health Organization growth standards [10]. The Nigerian sample includes 7361 (15.3% prevalence) wasted and 18,723 (38.9% prevalence) stunted children. The Kenyan sample includes 1775 (6.3% prevalence) wasted and 8396 (29.5% prevalence) stunted children. The dependent variables, wasting and stunting, are child-level composite binary indicators equal to one if the child’s calculated z-score is below negative two standard deviations from the reference median and zero otherwise.
Unique identifiers link the georeferenced data to records in the household surveys at the cluster level. However, the Demographic and Health Surveys employ geographic-masking with a coordinate displacement process to protect respondent confidentiality. The process displaces urban clusters up to two kilometers, displaces rural clusters up to five kilometers, and randomly selects one percent of the rural clusters to displace up to ten kilometers [71]. We link the Kids Recode files via timestamps and the cluster-level spatial identifiers to remotely monitored climatic variables (for more on variable composition, see Appendix A).
The Climate Hazards InfraRed Precipitation with Stations (CHIRPS) precipitation data product [75] and the Climate Hazards InfraRed Temperature with Stations (CHIRTS) temperature data product [76] provide interpolated high spatial resolution and coverage estimates of climatological activity between 1983 and 2016 inclusively. These products are used to estimate average total monthly rainfall, average total growing season rainfall, average maximum monthly temperature, and average maximum growing season temperature. And the normalized difference vegetation index (NDVI) is a continuous unitless greenness index of growing season abundance [77].
Selected covariates follow the UNICEF [29] conceptual framework along with spatially explicit temperature, precipitation, NDVI, and anomaly climatic inputs [75,76,77]. The emphasis of the model is on accommodating many possible determinants of malnutrition and prioritizing the most important within a specific contextual application while being easy to communicate across different users. Table 1, Table 2 and Table 3 report the summary statistics of discrete variables, continuous variables, and the hierarchical decomposition of DHS, respectively.

2.3. Statistical Analysis

Hierarchical models respect the heterogeneity of social experience and provides appropriate generalization to account for differences across groups [78]. Each child, household, cluster, and state has its own distinctive variation and characteristics. Hierarchical modeling attenuates the risk of both ecological and atomistic fallacies by considering all levels simultaneously [79].
Within a hierarchical modeling structure, we specify and measure the variability associated with different units of observation levels––child, household, cluster, and state––to match the Demographic and Health Surveys data structure. We assume each level is a pure hierarchical set, such that all clusters are contained within one and only one state, all households are contained within one and only one cluster, and all children are contained within one and only one household (for more on model specification, see Appendix B).
Variables at each level explain the measurement variability and its effect on malnutrition. Effects may also vary randomly among the units at higher levels (i.e., cross-level variability). For example, the magnitude of the effect of a child’s gender on their probability of being wasted may depend on cluster level characteristics, such as easy access to an improved toilet. Random variability may also exist at the household, cluster, or state scale––implying random intercepts. Explicit formulation of a hierarchical model with effects at, within, and between levels ameliorates these issues of conceptualization [80].
Inherently, child malnutrition is an individual and household-level phenomenon, yet it is at the country (and subnational) levels that many policy decisions are made. Using average data can be misleading if distribution is important and differs across countries and conclusions from cross-national data may not be applicable to individual countries’ situations [2]. Our approach addresses the interdependency explicitly. We preserve the units-of-analysis across levels in a combined structure and estimate random effects for each organizational unit. The standard error estimates incorporate the variability of the random effects and adjust for intraclass correlation [80].
Aggregation bias may occur when a variable has a different meaning, and thus a different effect, at different hierarchical levels. For example, the average quality of water and sanitation in a cluster may have an effect on a child’s health above and beyond the effect of an individual child’s water and sanitation circumstances at home. Our approach alleviates confounding effects by partitioning the effect of water and sanitation quality on health into separate components.
Misestimating precision may occur in standard error estimates if a model fails to account for dependence among individual responses within a group. Once the grouping has been established, even if it is established at random, the group itself will tend to become differentiated [81]. The group and its members can both influence and be influenced by the composition of the group [82]. Continuing from the previous example, the survey design may have selected the survey clusters at random yet the composition of children within a cluster is likely interdependent. An individual child’s water and sanitation circumstance is reliant on the available infrastructure and cultural conventions of that child’s community and so, too, is the child living next door, but far less so is the child living five states away.
We use the full assemblage of data across each organizational level to provide separate predicted probabilities for each category of interest. The estimators are weighted composites from the category of interest and the overall sample. Within group units are more similar than between group units and across levels, which mimics the first law of geography—everything is related to everything else, but near things are more related than distant things [83]. Children within the same household tend to be more similar to each other than those in other households, similarly for households within clusters, and clusters within states, and even for children within clusters, and households within states. This clustering occurs through some mechanism interconnected to unit characteristics: siblings do not end up in the same household by random chance. Generalization of classical regression methods with hierarchical methods is almost always an improvement in terms of fit, prediction and inference [84]. All results and conclusions are drawn from a four-level random intercept hierarchical generalized logit model specification.

3. Results

Figure 1 shows that the prevalence rates of wasting and stunting are overall spatially correlated, although there are pockets where rates deviate substantially suggesting different causal pathways [85]. The variable heterogeneity of malnutrition prevalence over the landscape highlights the need for a disaggregate and spatially explicit modeling approach (for more detailed spatial distributions of uncertainty estimates, see Appendix C). The Demographic and Health Surveys data form a natural hierarchical structure: regions within a country, states within a region, clusters within a state, households within a cluster, occupants within a household, and children for each woman.
The results across the various modeling approaches tell a consistent story, implying the results are robust to particular modeling variations. Our results as summarized in Table 4 indicate that the hierarchical structure alone explains 28 to 36 percent of the variation in malnutrition, meaning the additional model complexity has consequential explanatory value.
Figure 2 visualizes the results of the discrete covariates. It illustrates how much each categorical determinant affects malnutrition for a change from a baseline counterfactual. Generally, the effect sizes for stunting are larger than for wasting due to the smaller prevalence of wasting in the population. Because the model results measure the direct impact on the percentage point difference in probability of malnutrition in the population (i.e., prevalence), the size of the marginal effects have an upper-bound limit of the prevalence in the population. In other words, only already wasted and stunted children can transition to being non-wasted and non-stunted.

3.1. Social Determinants

Table 5 indicates that in both Nigeria and Kenya, the mother’s education plays a greater role in determining wasting, whereas household wealth is the leading determinant of stunting. On average, in Nigeria, the probability of being wasted is 4 percentage points (95% CI: −5.4 to −2.7) lower for a child from a mother with higher education than from a mother with no education, whereas in Kenya, the probability of being wasted is 1.7 percentage points (95% CI: −2.6 to −0.89) lower for a child from a mother with higher education than from a mother with no education. That is to say, the absolute prevalence rates of wasting in Nigeria and Kenya would drop from 15.31% and 6.25%, respectively, down to 11.31% and 4.55% if mothers of wasted children had higher education holding all else constant. Table 6 confirms that education plays a vital role in stunting prevalence as well. Mothers attaining higher education are associated with a reduction in stunting rates by 13 percentage points (95% CI: −16 to −10) in Nigeria and 5.9 percentage points (95% CI: −10 to −1.3) in Kenya. In other words, education alone has the potential to curtail the number of stunted children by over one-third.
In terms of quantifying the results in numbers of children, and in numbers of deaths prevented, the effects are highly epidemiologically significant. In 2011, the Nigeria under-five population was 27,195,000 with a 41% stunting prevalence (11,149,950) and a 14% wasting prevalence (3,807,300), with an overall mortality rate of 124/1000 for under-fives (3,372,180). This rate is much lower for non-malnourished children making the estimated projections of deaths prevented a conservative lower bound of the true value [10]. Using a maximally adjusted, minimum hazard ratio, of 2.12 for stunting mortality and 3.47 for wasting mortality, the mortality rate becomes 260/1000 at a minimum for stunted children, and 430/1000 at a minimum for wasted children [86]. At a maximum education is associated with a reduced stunting prevalence of 13 percentage points, or by 3,353,350 children, thus increasing education may prevent at a minimum 490,989 deaths. Similarly, if at a maximum education is associated with a reduced wasting prevalence of 4 percentage points, or by 1,087,800 children, increasing education may prevent at a minimum 315,767 deaths.
In 2011, the Kenya under-five population was 6,805,000 with a 35% stunting prevalence (2,381,750) and a 7% wasting prevalence (476,350), with an overall mortality rate of 73/1000 for under-fives (496,765). This rate is much lower for non-malnourished children, making the deaths prevented estimates conservative lower bounds of their true values [10]. Using a maximally adjusted minimum hazard ratio of 2.12 for stunting mortality and 3.47 for wasting mortality, the mortality rate becomes 150/1000 at a minimum for stunted children and 250/1000 at a minimum for wasted children [86]. If, at a maximum, education is associated with a reduced stunting prevalence of 5.9 percentage points, or by 401,495 children, then increasing education may prevent at a minimum 32,826 deaths. Similarly, if, at a maximum, education is associated with a reduced wasting prevalence of 1.7 percentage points, or by 115,685 children, increasing education may prevent at a minimum 21,206 deaths.
Results in Table 5 and Table 6 also confirm the powerful influence of wealth on malnutrition rates. The richest families from the highest wealth quintile exhibit a reduced wasting prevalence of 0.95 percentage points (95% CI: −2.5 to 0.63) in Nigeria and 1.2 percentage points (95% CI: −2.3 to −0.17) in Kenya. Wealth is associated with a reduced stunting prevalence of 16 percentage points (95% CI: −18 to −13 and −19 to −12) in both Nigeria and Kenya. The second highest wealth quintile is associated with a reduced wasting prevalence of 1.6 percentage points (95% CI: −2.8 to −0.42) in Nigeria and 1.1 percentage points (95% CI: −1.8 to −0.3) in Kenya. Similarly it is associated with a reduced stunting prevalence of 12 percentage points (95% CI: −15 to −9.9) in Nigeria and 10 percentage points (95% CI: −13 to −6.9) in Kenya (Table 5 and Table 6). Moving to the middle wealth quintile is associated with a reduced wasting prevalence of 1.3 percentage points (95% CI: −2.2 to −0.45) in Nigeria and 0.79 percentage points (95% CI: −1.5 to −0.04) in Kenya, and a reduced stunting prevalence of 6 percentage points (95% CI: −8.2 to −3.9) in Nigeria and 8.1 percentage points (95% CI: −11 to −5.4) in Kenya (Table 5 and Table 6). Even moving from the poorest to second-poorest wealth quintile is associated with a reduced wasting prevalence of 0.06 percentage points (95% CI: −0.92 to 0.8) in Nigeria and 0.92 percentage points (95% CI: −1.6 to −0.22) in Kenya, and a reduced stunting prevalence of 2.9 percentage points (95% CI: −4.7 to −1.1) in Nigeria and 4.3 percentage points (95% CI: −6.7 to −1.9) in Kenya. Overall, changes in wealth alone have a smaller but substantial impact on wasting with reductions up to one fifth. Even more substantially, wealth alone has the potential to curtail the number of stunted children by more than one half.

3.2. Environmental Determinants

The climate variables are a remote monitoring corollary to malnutrition. Climate variables have the potential to act as leading indicators for changes in malnutrition prevalence with wide coverage and lower costs compared to traditional clinical survey techniques. Malnutrition is often purported to be the most significant impact of climate change on children’s health, but little empirical evidence exists in the literature [36,87,88]. For Nigeria and Kenya, NDVI, precipitation, and temperature levels all play a significant, but not homogeneous, role in determining wasting and stunting prevalence.
Figure 3 shows that the increased precipitation levels in the preceding growing season may have an ameliorative effect. Precipitation levels in Kenya reaching an average 2.5 dm during the growing season is associated with a 3% predicted prevalence of wasting: an over 50% percent reduction from the sample average, whereas a precipitation level of 6.0 dm over the growing season in Nigeria is associated with a one in ten predicted prevalence of wasting. Figure 4 presents the effect of temperature on average predicted probability of malnutrition. In the case of Nigeria, it shows that higher temperatures correspond to higher wasting prevalence. On average, a maximum monthly temperature of 38 °C in the preceding growing season is associated with a 25% wasting prevalence. That is to say, the higher temperature corresponds to a 10-percentage point higher wasting prevalence. Under a forecasting regime, the results show that temperatures in Nigeria reaching an average monthly maximum of 38 °C during the growing season is associated with one in four children experiencing wasting the following year. Similarly, an average monthly maximum temperature of 35 °C in Kenya during the growing season is associated with one in ten children experiencing wasting the following: a nearly two-fold increase from the observed prevalence.
NDVI in the preceding growing season is a further measure with a strong inverse relationship to wasting rates. Figure 5 indicates that in both Nigeria and Kenya, moving from the lowest to the highest values of observable NDVI is associated with a reduced wasting rate of 50%.
While the absolute value or level plays the largest and most direct determining role in malnutrition outcomes, the long-term variability or anomaly plays a substantial secondary role, too. As presented in Table 5 and Table 6, one standard deviation increase in precipitation anomaly is associated with a reduced wasting prevalence of 0.12 percentage points (95% CI: −1.28 to 1.05) in Nigeria and a reduced stunting prevalence of 0.5 percentage points (95% CI: −1.18 to 0.18) in Kenya. Precipitation anomaly is associated with an increased wasting prevalence of 0.16 percentage points (95% CI: −0.13 to 0.46) in Kenya and an increased stunting prevalence of 1.36 percentage points (95% CI: −0.26 to 2.88) in Nigeria. One standard deviation increase in temperature anomaly is associated with a reduced wasting prevalence of 1.24 percentage points (95% CI: −2.39 to −0.12) in Nigeria and has zero detectable effect (95% CI: −0.28 to 0.27) in Kenya. Temperature anomaly is associated with a reduced stunting prevalence of 0.83 percentage points (95% CI: −2.67 to 0.97) in Nigeria, but an increased prevalence of 0.45 percentage points (95% CI: −0.19 to 1.08) in Kenya. One standard deviation increase in NDVI anomaly is associated with an increased wasting prevalence of 0.11 percentage points (95% CI: −0.36 to 0.6) in Nigeria and 0.19 percentage points (95% CI: −0.07 to 0.44) in Kenya. NDVI anomaly is associated with an increased stunting prevalence of 0.78 percentage points (95% CI: −0.52 to 2.08) in Nigeria, but a decreased prevalence of 0.44 percentage points (95% CI: −1.22 to 0.33) in Kenya.

3.3. Goodness of Fit

To assess fit we calculate percent correctly classified, McIntosh–Dorfman criterion, and McFadden’s pseudo-R-squared measurements (for specifics, see Appendix D). However, we rely on a decision analysis approach to generate predicted probability cutoff values, which we use to estimate the probability of underlying malnutrition. The decision curve analysis estimates of the hierarchical model specification are on average 15.3% for Nigeria wasting; 4.5% for Kenya wasting; 38.7% for Nigeria stunting; and 31.1% for Kenya stunting. The subsequent measures of sensitivity (true-positive rate) and specificity (true-negative rate) under the maximized net benefit regime range from 77.2% at a minimum to a maximum of 95.3% with a value of 84.2% on average, indicating a good fit.

3.4. Limitations

Using a standardized questionnaire model, the Demographic and Health Surveys Program aims to collect data that are comparable across countries. However, the questionnaire model has been modified across each of the seven phases of the Program making it difficult to measure changes through time. Given our samples are stratified and relatively large, sampling errors of excessive skepticism are low, but remain non-zero. Nor do we consider the overall quality of the anthropometric data itself as an approximation of normally distributed z-scores standard deviations [89].
In the survey design, individuals within households are not sampled, only clusters are sampled, and then households are sampled within clusters. Given that the DHS datasets do not provide a separate sampling fraction (i.e., weights) for clusters, households, and individuals for privacy, weighting in a multilevel model is infeasible [90,91,92]. The DHS geographic displacement process reduces the risk of disclosing confidential personal information, but adds artificial uncertainty into the signal-to-noise ratio and lowers the precision of estimated covariates.
Given the availability of DHS data release cycles coupled with the need to overlap with the spatially explicit historical CHIRTS and CHIRPS climate data, the temporal window for analysis is limited. And although the data have a temporal component, successive surveys are repeated cross sections, not a panel. There remains a need for similar studies that include more countries, across more surveys, across a broader timespan, examining more outcomes with more inputs specifically directed at the nexus of climate, conflict, and malnutrition.
Given the model specification complexity we did not control for potential misclassification error in the outcome variable, which may cause attenuated coefficient estimates [93,94]. Exploratory analysis suggests the accuracy of the observed wasted children is as low as 37% (Nigeria) and 21% (Kenya). The accuracy of the observed stunted children is better, 78% (Nigeria) and 66% (Kenya). Other estimates for the overdispersion of height-for-age z-scores suggest variance inflation factors as high as 110% [95].
Accompanying the model complexity is the opportunity to mis-specify the model. For example, it does not make sense to include a district-level random slope for the variable number of hospitals in a district as it does not vary within the district. However, low within-cluster variance is not much of an issue at all as long as there are an adequate number of clusters that do have sufficient data [96]. We performed specification robustness analyses, including linear probability specification and logit specifications (results available upon request). Given the discrete nature of the dependent variables, wasting and stunting, we use linear probability and logit models to motivate the initial coefficient interpretations and provide a lower bound on effect sizes. Exploring multiple model specifications helps to minimizes specification error and maximizes validity, and utilizing different populations defends against confounding [97].

4. Discussion

Although wasting and stunting are related malnutrition indicators their causal pathways, prevalence, duration, impact, and determining factors are distinct. Across both Nigeria and Kenya, stunting is most significantly impacted by family wealth, followed by the mother’s education, a clinical delivery, vaccinations, and children who are asymptomatic of fever, cough, or diarrhea. In Nigeria diet diversity manifests as a mitigating stunting risk factor, whereas in Kenya access to improved latrine facilities and rural households mitigates stunting prevalence. Wasting is most significantly impacted by mother’s education, followed by family wealth, a clinical delivery, and vaccinations across both Nigeria and Kenya. And in Nigeria, children living in urban households and those children exhibiting symptoms of fever, cough, or diarrhea are also at elevated wasting risk levels.
One likely cause of the discrepancy is the overall sample dispersion across the two countries, coupled with the greater potential reduction for Nigeria to curb malnutrition prevalence given its higher observed rates. For example, in Nigeria, mothers’ education exhibits a more bimodal distribution, where over 45% of mothers have no formal education and over 30% have secondary or higher, whereas in Kenya, there is a more clumped distribution centered around primary education, which accounts for over 54% of the sample. Given these two very different distributions, the model results suggest that in a more divided society with a high prevalence of malnutrition, having greater access to education will afford proportionally greater gains in beneficial public health outcomes.
Climatic variables are powerful determinants of malnutrition. Across the observable range of values, changes in precipitation, temperature, or NDVI (in the preceding growing season) alone could curtail or inflate the number of wasted and stunted children by more than one half. However, their effects can vary greatly across different nutrition indicators and different countries (congruent with previous efforts elsewhere) [42,98]. Due to the distinct causal pathways and chronic nature of stunting, the signal-to-noise ratio of climate determinants is markedly diminished. In Kenya, higher precipitation and NDVI levels were deleterious and significant determinants, while higher temperature levels were a mitigating and significant determinant. Yet, in Nigeria higher temperature, precipitation, and NDVI levels were mitigating determinants.
Surprisingly, some oft-purported determinants of malnutrition were not significant in either the statistical or epidemiological sense. These include climate anomalies, access to improved latrine facilities, access to improved water facilities, weaning practices, and diet diversity for wasting. Similarly, for stunting, improved water facilities and weaning are not significant. Further research is needed to ameliorate these discrepancies. In particular, agencies and organizations such as the Famine Early Warning Systems Network, Action Against Hunger, the International Food Policy Research Institute, the World Bank, the World Food Program, UNICEF, the World Health Organization, and many others rely accurate and precise information to inform their models, forecasts, and resource deployment.
One segment of the public health community has decided that certain determinants are a public health boon. Therefore, these determinants must be drastically buoyed. The force with which these conclusions are presented is not in reasonable balance with the strength of the evidence. Programs, once in place develop a life of their own; the possibility of health benefits becomes probability, which becomes certainty. The appearance of scientific unanimity is a powerful political tool, especially when the evidence is mixed. Dissent becomes a threat, which must be marginalized. If funding agencies and journals are unwilling to brook opposition, rational discussion is curtailed. There soon comes about the pretense of policy based on scientific inquiry—without the substance.
In the clinical setting of public health and epidemiology, a diagnostic application helps to estimate the probability that malnutrition is present, identifying the nature or cause of the malnutrition; whereas a prognostic application helps to predict how malnutrition will develop and target preventive interventions to children at relatively high risk [99]. Diagnostics can be described as the probability of malnutrition conditional on a set of latent determinants, whereas prognostics can be thought of as the obverse or the probability of future outcomes conditional on being malnourished.
Estimated effects from latent determinants provide the diagnostic insights, whereas harm versus benefit establishes the prognostic framework. The purpose of a prognostic model is that better decisions are made with the model than without. Within the prognostic framework reliability of predictions is key. It is our aim to estimate the probability of malnutrition in a diagnostic sense, and to help target preventive interventions in a prognostic sense. Understanding how the distinctiveness of location effects malnutrition provides even more clarity.

5. Conclusions

Malnutrition devastates millions of children every year, yet the latent determinants are largely obscure. One is best informed by examining determinants on the basis of quantitative and epidemiological significance. A determinant’s impact is best measured by its ability to change malnutrition prevalence in an epidemiologically significant way. We find the most impactful latent determinants each have the capacity to reduce prevalence rates by as much as 50%: an epidemiologically significant effect.
The inconsistencies of determinants across space and malnutrition outcomes highlight the need for prudent, highly specific, and tailored approaches, especially when using climate determinants for any forecasting efforts or policy interventions [100,101]. Particular focus should be paid to those determinants that are either actionable by policy intervention or serviceable in forecasting and intervention efforts as well as epidemiologically significant. Identifying effective mitigating determinants to prevent the harmful effects of malnutrition in children should be a priority. Only with explicit identification and measurement can intervention organizations and governments begin to make substantial progress to reduce childhood malnutrition.

Author Contributions

Conceptualization, A.S. and L.S.; methodology, A.S. and L.S.; software, A.S.; validation, A.S.; formal analysis, A.S.; investigation, A.S. and L.S.; resources, A.S. and L.S.; data curation, A.S.; writing—original draft preparation, A.S.; writing—review and editing, A.S. and L.S.; visualization, A.S.; supervision, L.S.; project administration, L.S.; funding acquisition, L.S. All authors have read and agreed to the published version of the manuscript.

Funding

Initial research was partially supported by the by the UK Aid from the UK Department for International Development, in collaboration with Action Against Hunger, on behalf of the Modelling Early Risk Indicators to Anticipate Malnutrition consortium (Project Identifier: GB-CHC-1047501-MERIAM). The funder of this study had no role in the study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had final responsibility for the decision to submit for publication.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

No new data were created for this study. Restrictions apply to the availability of some of these data. Data were obtained from The Demographic and Health Surveys available at https://www.dhsprogram.com/Data/ (accessed on 25 September 2017) and from the Climate Hazards Center available at https://www.chc.ucsb.edu/data (accessed on 25 September 2017) with the permission of the providers.

Acknowledgments

The authors thank Emily Dacquisto, David Baker, Honglin Zhong, Peter White, Molly Brown, Paul K. Huth, and Klaus Hubacek for their comments, suggestions, and technical assistances across various stages of this research.

Conflicts of Interest

The authors declare no conflicts of interest.

Appendix A. Variable Composition

Appendix A.1. Child-Level

The sex variable is an unadulterated binary indicator equal to one if the child is female and zero if male. The delivery variable is a collapsed binary indicator equal to one if delivery occurred in a hospital facility or health clinic and zero otherwise. The birth variable is a collapsed binary indicator equal to one if the delivery was a singleton birth and zero if delivery involved multiple births (i.e., twins). The weaned variable is a composite categorical indicator equal to one if the child is weaned by 1 year, two if the child is breastfed up to 1 year, three if the child is weaned before 1 year, and zero if the child is breastfed beyond 1 year. The vaccines—minimum variable is a composite binary indicator equal to one if the child received at least 1 of 9 vaccinations (Polio 0, 1, 2, 3; DPT 1, 2, 3; BCG, and Measles) and zero otherwise. The vaccines—maximum variable is a composite binary indicator equal to one if the child received all nine vaccinations and zero otherwise. The diet variable is a composite binary indicator equal to one if the child received a diverse variety of 4 or more food groups (of seven possible, including grains, legumes, dairy, meat or fish, eggs, fruits and vegetables high in vitamin A, and other fruits and vegetables) in the past 24 h, or three or more food groups plus breast milk, and zero otherwise. The sick variable is a composite binary indicator equal to one if the child is asymptomatic (i.e., did not present with diarrhea, a fever, or a cough in the past 2 weeks) and zero otherwise. The child’s age variable is an unadulterated continuous indicator of the child’s age in months, from date of birth to date of interview.

Appendix A.2. Household-Level

The latrine variable is a composite binary indicator equal to one if the facility is “improved”, meaning it is not shared and the type of toilet facility for the household is a flush toilet (either to a sewer system, septic tank, pit, or anywhere else); ventilated improved pit latrine or pit latrine with slab; or a composting toilet and zero otherwise if the facility is shared or an open pit; no facility, brush or field; bucket toilet; hanging toilet; or anywhere else. The water variable is a collapsed binary indicator equal to one if it is “improved”, meaning the major source of drinking water for the household is piped water into the dwelling, yard, or plot; a public tap, standpipe, or borehole; a protected well or protected spring water; rainwater; or bottled water and zero otherwise from sources including unprotected wells or springs, water delivered by tanker trucks, or surface water. The mother’s education variable is an unadulterated standardized categorical indicator of highest education level attended equal to zero if no education, one if primary, two if secondary, and three if higher. The wealth index variable is an unadulterated composite categorical indicator of a household’s cumulative standard of living, calculated using ownership of assets (e.g., televisions and bicycles); housing construction materials; types of water and latrine facilities, and generated by placing all interviewed households along a continuous scale of relative wealth and then separating them into five wealth quintiles: poorest, poorer, middle, richer, and richest. The mother’s age variable is an unadulterated continuous indicator of the mother’s current age in completed years of decades, from date of birth to date of interview. The birth tally variable is an unadulterated continuous indicator of the total number of births of the mother.

Appendix A.3. Cluster-Level

The residence variable is an unadulterated binary indicator equal to one if the de facto place of residence is rural (based on whether the cluster is defined as rural in the sample design––a country-specific designation) and zero if defined as urban. The precipitation variable is a composite continuous measure of average total monthly rainfall in decimeters during the preceding growing season derived from the Climate Hazards Group InfraRed Precipitation with Station (CHIRPS) dataset replete with a 0.05° spatial resolution [75]. The temperature variable is a composite continuous measure of average maximum monthly temperature in Celsius during the preceding growing season derived from the Climate Hazards Group InfraRed Temperature with Station (CHIRTS) dataset replete with a 0.05° spatial resolution [75]. The precipitation anomaly variable is a composite continuous measure of average monthly rainfall anomaly from the previous five-year average in decimeters during the preceding growing season. The temperature anomaly variable is a composite continuous measure of average maximum monthly temperature anomaly from the previous five-year average in Celsius during the preceding growing season. The greenness index variable is a composite continuous unit-less index measure between zero and one of the normalized difference vegetation index (NDVI) for the three greenest months during the preceding growing season replete with a 0.05° spatial resolution [77].

Appendix A.4. State-Level and Other Controls

State-level indicators are First-level Administrative Divisions and include 47 counties for Kenya and 36 states plus one federal capital territory for Nigeria. The interview month variable is an unadulterated categorical control indicator of the month in which the survey took place. The survey phase variable is an unadulterated categorical control indicator of the phase in which the survey took place (DHS-IV from 1997 to 2003; DHS-V from 2003 to 2008; and DHS-VI from 2008 to 2013) [90,91,92].

Appendix B. Hierarchical Specification

The unconditional, intercept-only model assumes random-effects coefficients have a mean of zero and determines the variance components. The variance decomposition shows both the within- and between-group variability for the proportion of the variance of the outcome. This procedure provides evidence to justify the application of a hierarchical model.
Child-level:
Y i j k l = β 0 j k l + e i j k l
Y i j k l : anthropometric indicator (e.g., wasting or stunting) of child i , in household j , in cluster k , in state l .
β 0 j k l : mean indicator of household j , in cluster k , in state l .
e i j k l : random child effect, deviation of child i j k l ’s indicator from household j k l ’s mean; ~ N ( 0 , σ 2 ) .
Household-level:
β 0 j k l = θ 00 k l + r 0 j k l
θ 00 k l : mean indicator of cluster k , in state l .
r 0 j k l : random household effect, deviation of household j k l ’s indicator mean from cluster k l ’s mean; ~ N ( 0 , τ β )
Cluster-level:
θ 00 k l = η 000 l + v 00 k l
η 000 l : mean indicator of state l .
v 00 k l : random cluster effect, deviation of cluster k l ’s indicator mean from state l ’s mean; ~ N ( 0 , τ θ ) .
State-level:
η 000 l = γ 0000 + u 000 l
γ 0000 : grand indicator mean.
u 000 l : random state effect, deviation of state l ’s indicator mean from grand mean; ~ N ( 0 , τ η ) .
The subscripts i , j , k , and l denote children, households, clusters, and states where,
i = 1 ,   2 , … ,   N j k l children within household j , in cluster k , in state l ;
j = 1 ,   2 , … ,   J k l households, within cluster k , in state l ;
k = 1 ,   2 , … ,   K l clusters, within state l ; and
l = 1 ,   2 , … ,   L states.
The complementary mixed model for the unconditional, intercept-only model is given by,
Y i j k l = γ 0000 + u 000 l + v 00 k l + r 0 j k l + e i j k l .
The unconditional, intercept-only specification estimates the grand mean, γ 0000 too, and provides information about the outcome variability at each level. The proportion of variance in the outcome variable explained by the grouping structure of the hierarchical model is given by
Var ( Y i j ) = Var ( u 000 l + v 00 k l + r 0 j k l + e i j k l ) = τ η + τ θ + τ β + σ 2 .
Total variability in outcome Y i j k l is partitioned across each level: level 1, σ 2 among children within households; level 2, τ β among households within clusters; level 3, τ θ among clusters within states; and level 4, τ η among sates. The proportion of variation attributed to each level is given by
Level 1:
  σ 2 τ η + τ θ + τ β + σ 2
Level 2:
  τ β τ η + τ θ + τ β + σ 2
Level 3:
  τ θ τ η + τ θ + τ β + σ 2
Level 4:
  τ η τ η + τ θ + τ β + σ 2
The intraclass correlation coefficient or the cluster effect is denoted by the coefficient rho. For the four-level fully unconditional hierarchical model, the intraclass correlation coefficients are given by
Level 2:
  ρ 2 = τ η + τ θ + τ β τ η + τ θ + τ β + σ 2
Level 3:
ρ 3 = τ η + τ θ τ η + τ θ + τ β + σ 2
Level 4:
ρ 4 = τ η τ η + τ θ + τ β + σ 2
The level-1 intraclass correlation is undefined, and by definition, ρ 2 ≥ ρ 3 ≥ ρ 4 .
The second-order specification for the four-level hierarchical model permits random intercepts, which account for the unique effects of each household, cluster, and state on the anthropometric indicator outcome variable.
Child-level:
Y i j k l = β 0 j k l + ∑ p = 1 P β p j k l X .1 p i j k l + e i j k l
Y i j k l : anthropometric indicator of child i , in household j , in cluster k , in state l .
β 0 j k l : child-level intercept for household j , in cluster k , in state l , which varies between children according to the household-level specification.
β p j k l : child-level fixed effects coefficients for each child-level characteristic X .1 p i j k l .
X .1 p i j k l : p = 1 , … P child-level characteristics.
e i j k l : random child effect, deviation of child i j k l ’s indicator from the predicted indicator; ~ N ( 0 , σ 2 ) .
Household-level:
β 0 j k l = θ 00 k l + ∑ q = 1 Q θ 0 q k l X .2 q j k l + r 0 j k l  
β p j k l = θ p 0 k l   ∀   p
θ 00 k l : household-level intercept for cluster k , in state l , which varies between households according to the cluster-level specification.
θ 0 q k l : household-level fixed effects coefficients for each household-level characteristic X .2 q j k l .
θ p 0 k l : equivalent child-level fixed effects in household-level notation.
X .2 q j k l : q = 1 , … , Q household-level characteristics.
r 0 j k l : random household effect, deviation of household j k l ’s indicator from the predicted indicator; ~ N ( 0 , τ β ) .
Cluster-level:
θ 00 k l = η 000 l + ∑ r = 1 R η 00 r l X .3 r k l + v 00 k l  
θ 0 q k l = η 0 q 0 l   ∀   q
θ p 0 k l = η p 00 l   ∀   p
η 000 l : cluster-level intercept for state l , which varies between clusters according to the state-level specification.
η 00 r l : cluster-level fixed effects coefficients for each cluster-level characteristic X .3 r k l .
η 0 q 0 l : equivalent household-level fixed effects in cluster-level notation.
η p 00 l : equivalent child-level fixed effects in cluster-level notation.
X .3 r k l : r = 1 , … , R cluster-level characteristics.
v 00 k l : random cluster effect, deviation of cluster k l ’s indicator from the predicted indicator; ~ N ( 0 , τ θ ) .
State-level:
η 000 l = γ 0000 + ∑ s = 1 S γ 000 s X .4 s l + u 000 l  
η 00 r l = γ 00 r 0   ∀   r
η 0 q 0 l = γ 0 q 00   ∀   q
η p 00 l = γ p 000   ∀   p
γ 0000 : grand intercept.
γ 000 s : state-level fixed-effects coefficients for each state-level characteristic X .4 s l .
γ 00 r 0 : equivalent cluster-level fixed effects in state-level notation.
γ 0 q 00 : equivalent household-level fixed effects in state-level notation.
γ p 000 : equivalent child-level fixed effects in state-level notation.
X .4 s l : s = 1 , … S state-level characteristics.
u 000 l : random state effect, deviation of state l ’s indicator from the predicted indicator; ~ N ( 0 , τ η ) .
The mixed model is given by
Y i j k l = γ 0000 + ∑ p = 1 P γ p 000 X .1 p i j k l + ∑ q = 1 Q γ 0 q 00 X .2 q j k l + ∑ r = 1 R γ 00 r 0 X .3 r k l       + ∑ s = 1 S γ 000 s X .4 s l + u 000 l + v 00 k l + r 0 j k l + e i j k l
where it is easier to parse the model composition in terms of γ representing the fixed effects and u 000 l , v 00 k l , r 0 j k l , and e i j k l representing the random effects.

Appendix C. Spatial Dispersions

Nutrients 16 02014 g0a1
Figure A1. Empirical Bayesian kriging model uncertainty estimates across Kenya and Nigeria based on projections in Figure 1.
Figure A1. Empirical Bayesian kriging model uncertainty estimates across Kenya and Nigeria based on projections in Figure 1.
Nutrients 16 02014 g0a1

Appendix D. Full Model Results

Table A1. Hierarchical results: wasted.
Table A1. Hierarchical results: wasted.
Wasted
Hierarchical Random InterceptNigeriaKenya
Average Marginal Effects with 95% Confidence Interval in Brackets
Sex—Female−0.012 ***[−0.019, −0.0049]−0.0075 ***[−0.011, −0.0036]
Delivery—Clinic−0.0091 **[−0.017, −0.0011]−0.010 ***[−0.016, −0.0046]
Birth—Singleton−0.041 ***[−0.067, −0.014]−0.032 ***[−0.055, −0.010]
Weaned—By 1 Year Old−0.0044 [−0.012, 0.0034]−0.0011 [−0.0048, 0.0026]
Vaccines—Minimum−0.010 **[−0.020, −0.00032]−0.0044 [−0.014, 0.0052]
Vaccines—Maximum−0.010 *[−0.020, 0.000018]−0.0027 [−0.0085, 0.0032]
Diet—Diverse0.0077 *[−0.00096, 0.016]−0.0032 [−0.0090, 0.0025]
Sick—Asymptomatic−0.010 ***[−0.018, −0.0025]−0.0016 [−0.0058, 0.0026]
Latrine—Improved−0.0031 [−0.010, 0.0038]0.0045 [−0.0038, 0.013]
Water—Improved−0.0026 [−0.012, 0.0066]−0.00021 [−0.0041, 0.0037]
Residence—Rural−0.0086 [−0.022, 0.0047]−0.00029 [−0.0051, 0.0046]
Mother’s Education
 Primary−0.0096 **[−0.018, −0.0012]−0.011 ***[−0.017, −0.0056]
 Secondary−0.020 ***[−0.028, −0.011]−0.0089 **[−0.016, −0.0018]
 Higher−0.040 ***[−0.054, −0.027]−0.017 ***[−0.026, −0.0089]
Wealth Index
 Poorer−0.00059 [−0.0092, 0.0080]−0.0092 ***[−0.016, −0.0022]
 Middle−0.013 ***[−0.022, −0.0045]−0.0079 **[−0.015, −0.00044]
 Richer−0.016 ***[−0.028, −0.0042]−0.011 ***[−0.018, −0.0030]
 Richest−0.0095 [−0.025, 0.0063]−0.012 **[−0.023, −0.0017]
Child’s Age−0.022 ***[−0.028, −0.015]−0.0013 [−0.0038, 0.0012]
Mother’s Age0.0026 [−0.0064, 0.012]−0.0022 [−0.0065, 0.0020]
Birth Tally−0.0017 [−0.0039, 0.00054]0.00069 [−0.00072, 0.0021]
NDVI−0.092 ***[−0.14, −0.049]−0.039 ***[−0.066, −0.013]
NDVI Anomaly0.044 [−0.14, 0.23]0.055 [−0.021, 0.13]
Precipitation−0.0096 [−0.023, 0.0041]−0.015 ***[−0.025, −0.0063]
Precipitation Anomaly−0.0045 [−0.049, 0.040]0.011 [−0.0088, 0.031]
Temperature0.012 ***[0.0079, 0.015]0.0024 ***[0.0012, 0.0036]
Temperature Anomaly−0.027 **[−0.052, −0.0026]−0.000052 [−0.0062, 0.0061]
Fixed Effect—Month and PhaseYes Yes
Number of Observations44,717 26,130
Log Pseudo Likelihood−17,439.97−5572.63
Predicted Outcome AnalysisStandardMax Net BenefitStandardMax Net Benefit
McIntosh–Dorfman Criterion1.171.711.021.73
Percent Correctly Classified86.8885.7693.7879.93
Sensitivity17.9285.481.5294.36
Specificity99.5585.81100.0078.96
Net Benefit0.0270.1110.0010.046
Cut-Off Value0.50.1580.50.045
* p < 0.10, ** p < 0.05, *** p < 0.01.
Table A2. ICC and variance decomposition: wasted.
Table A2. ICC and variance decomposition: wasted.
Wasted
Hierarchical Random InterceptNigeriaKenya
Random Effect—Variance Component with 95% Confidence Interval in Brackets
States0.3[0.18, 0.41]0.37[0.18, 0.57]
Clusters0.47[0.31, 0.62]0.14[0.025, 0.25]
Households1.17[0.86, 1.47]1.19[0.70, 1.67]
Intraclass Correlation—Coefficients with 95% Confidence Interval in Brackets
States0.057[0.039, 0.081]0.075[0.044, 0.124]
Clusters0.146[0.114, 0.186]0.103[0.07, 0.149]
Households0.370[0.318, 0.425]0.340[0.277, 0.41]
Variance Decomposition—Percent by Level
States5.66% 7.50%
Clusters8.96% 2.77%
Households22.34% 23.76%
Children63.04% 65.96%
Table A3. Hierarchical results: stunted.
Table A3. Hierarchical results: stunted.
Stunted
Hierarchical Random InterceptNigeriaKenya
Average Marginal Effects with 95% Confidence Interval in Brackets
Sex—Female−0.051 ***[−0.059, −0.042]−0.077 ***[−0.091, −0.063]
Delivery—Clinic−0.022 ***[−0.033, −0.011]−0.046 ***[−0.064, −0.028]
Birth—Singleton−0.13 ***[−0.17, −0.091]−0.23 ***[−0.28, −0.18]
Weaned—By 1 Year Old−0.0031 [−0.016, 0.0100]−0.011 [−0.028, 0.0065]
Vaccines—Minimum−0.0056 [−0.028, 0.017]−0.029 **[−0.055, −0.0020]
Vaccines—Maximum−0.040 ***[−0.060, −0.019]−0.016 **[−0.031, −0.0022]
Diet—Diverse−0.020 **[−0.036, −0.0037]−0.0051 [−0.023, 0.013]
Sick—Asymptomatic−0.034 ***[−0.050, −0.018]−0.013 **[−0.026, −0.00072]
Latrine—Improved−0.0043 [−0.020, 0.011]−0.050 ***[−0.072, −0.028]
Water—Improved0.0020 [−0.011, 0.015]−0.011 [−0.027, 0.0051]
Residence—Rural0.015 **[0.000097, 0.029]−0.014 [−0.036, 0.0080]
Mother’s Education
 Primary−0.015 **[−0.030, −0.000055]0.025 [−0.0050, 0.056]
 Secondary−0.054 ***[−0.074, −0.034]−0.025 [−0.057, 0.0084]
 Higher−0.13 ***[−0.16, −0.10]−0.059 **[−0.10, −0.013]
Wealth Index
 Poorer−0.029 ***[−0.047, −0.011]−0.043 ***[−0.067, −0.019]
 Middle−0.060 ***[−0.082, −0.039]−0.081 ***[−0.11, −0.054]
 Richer−0.12 ***[−0.15, −0.099]−0.10 ***[−0.13, −0.069]
 Richest−0.16 ***[−0.18, −0.13]−0.16 ***[−0.19, −0.12]
Child’s Age−0.0075 [−0.017, 0.0016]−0.026 ***[−0.033, −0.018]
Mother’s Age−0.036 ***[−0.047, −0.025]−0.046 ***[−0.061, −0.031]
Birth Tally0.0037 **[0.00062, 0.0068]0.011 ***[0.0067, 0.016]
NDVI−0.066 [−0.19, 0.061]0.12 ***[0.057, 0.18]
NDVI Anomaly0.30 [−0.20, 0.80]−0.13 [−0.36, 0.098]
Precipitation−0.015 [−0.044, 0.014]0.033 **[0.0033, 0.063]
Precipitation Anomaly0.052 [−0.010, 0.11]−0.034 [−0.080, 0.012]
Temperature−0.0026 [−0.013, 0.0073]−0.0092 ***[−0.012, −0.0061]
Temperature Anomaly−0.018 [−0.058, 0.021]0.010 [−0.0043, 0.024]
Fixed Effect—Month and PhaseYes Yes
Number of Observations44,717 26,130
Log Pseudo-Likelihood−26,250.40−14,400.96
Predicted Outcome AnalysisStandardMax Net BenefitStandardMax Net Benefit
McIntosh–Dorfman Criterion1.561.601.431.70
Percent Correctly Classified80.2879.2082.4985.92
Sensitivity66.3482.3244.3182.56
Specificity89.2477.2098.2987.31
Net Benefit0.2170.2330.1250.205
Cut-Off Value0.50.3830.50.317
** p < 0.05, *** p < 0.01.
Table A4. ICC and variance decomposition: stunted.
Table A4. ICC and variance decomposition: stunted.
Stunted
Hierarchical Random InterceptNigeriaKenya
Random Effect—Variance Component with 95% Confidence Interval in Brackets
States0.26[0.16, 0.35]0.099[0.048, 0.15]
Clusters0.22[0.17, 0.26]0.13[0.071, 0.20]
Households0.81[0.69, 0.93]1.16[0.89, 1.43]
Intraclass Correlation—Coefficients with 95% Confidence Interval in Brackets
States0.056[0.039, 0.08]0.021[0.013, 0.035]
Clusters0.103[0.083, 0.127]0.050[0.037, 0.068]
Households0.281[0.257, 0.306]0.297[0.257, 0.341]
Variance Decomposition—Percent by Level
States5.58% 2.12%
Clusters4.74% 2.87%
Households17.74% 24.71%
Children5.58% 2.12%

References

  1. Black, R.E.; Victora, C.G.; Walker, S.P.; Bhutta, Z.A.; Christian, P.; de Onis, M.; Ezzati, M.; Grantham-McGregor, S.; Katz, J.; Martorell, R.; et al. Maternal and child undernutrition and overweight in low-income and middle-income countries. Lancet 2013, 382, 427–451. [Google Scholar] [CrossRef] [PubMed]
  2. Smith, L.C.; Haddad, L.J. Explaining Child Malnutrition in Developing Countries: A Cross-Country Analysis; International Food Policy Research Institute: Washington, DC, USA, 2000. [Google Scholar]
  3. de Onis, M.; Branca, F. Childhood stunting: A global perspective. Matern. Child Nutr. 2016, 12, 12–26. [Google Scholar] [CrossRef] [PubMed]
  4. WHO. Global Nutrition Targets 2025: Policy Brief Series; WHO: Geneva, Switzerland, 2014.
  5. Nations, U. Transforming Our World: The 2030 Agenda for Sustainable Development; United Nations: New York, NY, USA, 2015.
  6. UNICEF. Progress for Every Child in the SDG Era; UNICEF: New York, NY, USA, 2018. [Google Scholar]
  7. Tzioumis, E.; Adair, L.S. Childhood dual burden of under- and overnutrition in low- and middle-income countries: A critical review. Food Nutr. Bull. 2014, 35, 230–243. [Google Scholar] [CrossRef] [PubMed]
  8. Mehta, N.M.; Corkins, M.R.; Lyman, B.; Malone, A.; Goday, P.S.; Carney, L.; Monczka, J.L.; Plogsted, S.W.; Schwenk, W.F.; American Society for Parenteral and Enteral Nutrition (ASPEN) Board of Directors. Defining pediatric malnutrition: A paradigm shift toward etiology-related definitions. J. Parenter. Enter. Nutr. 2013, 37, 460–481. [Google Scholar] [CrossRef] [PubMed]
  9. Sandler, A.M. The legacy of a standard of normality in child nutrition research. SSM-Popul. Health 2021, 15, 100865. [Google Scholar] [CrossRef] [PubMed]
  10. UNICEF. Improving Child Nutrition: The Achievable Imperative for Global Progress; UNICEF: New York, NY, USA, 2013; pp. 1–114. [Google Scholar]
  11. Akombi, B.J.; Agho, K.E.; Hall, J.J.; Merom, D.; Astell-Burt, T.; Renzaho, A.M. Stunting and severe stunting among children under-5 years in Nigeria: A multilevel analysis. BMC Pediatr. 2017, 17, 15. [Google Scholar] [CrossRef] [PubMed]
  12. Akombi, B.J.; Agho, K.E.; Merom, D.; Hall, J.J.; Renzaho, A.M. Multilevel Analysis of Factors Associated with Wasting and Underweight among Children Under-Five Years in Nigeria. Nutrients 2017, 9, 17. [Google Scholar] [CrossRef] [PubMed]
  13. Black, R.E.; Allen, L.H.; Bhutta, Z.A.; Caulfield, L.E.; de Onis, M.; Ezzati, M.; Mathers, C.; Rivera, J.; Group, M.a.C.U.S. Maternal and child undernutrition: Global and regional exposures and health consequences. Lancet 2008, 371, 243–260. [Google Scholar] [CrossRef] [PubMed]
  14. Boah, M.; Azupogo, F.; Amporfro, D.A.; Abada, L.A. The epidemiology of undernutrition and its determinants in children under five years in Ghana. PLoS ONE 2019, 14, e0219665. [Google Scholar] [CrossRef] [PubMed]
  15. Brown, M.E. Famine Early Warning Systems and Remote Sensing Data; Springer Science & Business Media: Berlin/Heidelberg, Germany, 2008. [Google Scholar]
  16. Darteh, E.K.M.; Acquah, E.; Kumi-Kyereme, A. Correlates of stunting among children in Ghana. BMC Public Health 2014, 14, 504. [Google Scholar] [CrossRef] [PubMed]
  17. de Groot, R.; Palermo, T.; Handa, S.; Ragno, L.P.; Peterman, A. Cash Transfers and Child Nutrition: Pathways and Impacts. Dev. Policy Rev. 2017, 35, 621–643. [Google Scholar] [CrossRef] [PubMed]
  18. Engebretsen, I.M.S.; Tylleskär, T.; Wamani, H.; Karamagi, C.; Tumwine, J.K. Determinants of infant growth in Eastern Uganda: A community-based cross-sectional study. BMC Public Health 2008, 8, 418. [Google Scholar] [CrossRef] [PubMed]
  19. Engle, P.L.; Menon, P.; Haddad, L. Care and nutrition: Concepts and measurement. World Dev. 1999, 27, 1309–1337. [Google Scholar] [CrossRef]
  20. Fernandes, E.C.B.; Castro, T.G.D.; Sartorelli, D.S. Associated factors of malnutrition among African children under five years old, Bom Jesus, Angola. Rev. Nutr. 2017, 30, 33–44. [Google Scholar] [CrossRef]
  21. Fernandez, I.D.; Himes, J.H.; de Onis, M. Prevalence of nutritional wasting in populations: Building explanatory models using secondary data. Bull. World Health Organ. 2002, 80, 282–291. [Google Scholar] [PubMed]
  22. Habaasa, G. An investigation on factors associated with malnutrition among underfive children in Nakaseke and Nakasongola districts, Uganda. BMC Pediatr. 2015, 15, 134. [Google Scholar] [CrossRef] [PubMed]
  23. Kavle, J.A.; El-Zanaty, F.; Landry, M.; Galloway, R. The rise in stunting in relation to avian influenza and food consumption patterns in Lower Egypt in comparison to Upper Egypt: Results from 2005 and 2008 Demographic and Health Surveys. BMC Public Health 2015, 15, 285. [Google Scholar] [CrossRef] [PubMed]
  24. Lesiapeto, M.S.; Smuts, C.M.; Hanekom, S.M.; Du Plessis, J.; Faber, M. Risk factors of poor anthropometric status in children under five years of age living in rural districts of the Eastern Cape and KwaZulu-Natal provinces, South Africa. S. Afr. J. Clin. Nutr. 2010, 23, 202–207. [Google Scholar] [CrossRef]
  25. Müller, O.; Krawinkel, M. Malnutrition and health in developing countries. CMAJ 2005, 173, 279–286. [Google Scholar] [CrossRef] [PubMed]
  26. Ricci, C.; Carboo, J.; Asare, H.; Smuts, C.M.; Dolman, R.; Lombard, M. Nutritional status as a central determinant of child mortality in sub-Saharan Africa: A quantitative conceptual framework. Matern. Child Nutr. 2019, 15, e12722. [Google Scholar] [CrossRef] [PubMed]
  27. Smith, L.C.; Haddad, L. Reducing child undernutrition: Past drivers and priorities for the post-MDG era. World Dev. 2015, 68, 180–204. [Google Scholar] [CrossRef]
  28. Stewart, C.P.; Iannotti, L.; Dewey, K.G.; Michaelsen, K.F.; Onyango, A.W. Contextualising complementary feeding in a broader framework for stunting prevention. Matern. Child Nutr. 2013, 9, 27–45. [Google Scholar] [CrossRef] [PubMed]
  29. UNICEF. The State of the World’s Children; Oxford University Press: Oxford, UK, 1998. [Google Scholar]
  30. Wamani, H.; Åstrøm, A.N.; Peterson, S.; Tumwine, J.K.; Tylleskär, T. Predictors of poor anthropometric status among children under 2 years of age in rural Uganda. Public Health Nutr. 2006, 9, 320–326. [Google Scholar] [CrossRef] [PubMed]
  31. Willey, B.A.; Cameron, N.; Norris, S.A.; Pettifor, J.M.; Griffiths, P.L. Socio-economic predictors of stunting in preschool children–a population-based study from Johannesburg and Soweto. S. Afr. Med. J. 2009, 99, 450–456. [Google Scholar] [PubMed]
  32. Bhutta, Z.A.; Ahmed, T.; Black, R.E.; Cousens, S.; Dewey, K.; Giugliani, E.; Haider, B.A.; Kirkwood, B.; Morris, S.S.; Sachdev, H.P.S.; et al. What works? Interventions for maternal and child undernutrition and survival. Lancet 2008, 371, 417–440. [Google Scholar] [CrossRef] [PubMed]
  33. Keino, S.; Plasqui, G.; Ettyang, G.; van den Borne, B. Determinants of stunting and overweight among young children and adolescents in sub-Saharan Africa. Food Nutr. Bull. 2014, 35, 167–178. [Google Scholar] [CrossRef] [PubMed]
  34. Brown, M.E.; Backer, D.; Billing, T.; White, P.; Grace, K.; Doocy, S.; Huth, P. Empirical studies of factors associated with child malnutrition: Highlighting the evidence about climate and conflict shocks. Food Secur. 2020, 12, 1241–1252. [Google Scholar] [CrossRef]
  35. Akombi, B.J.; Agho, K.E.; Hall, J.J.; Wali, N.; Renzaho, A.; Merom, D. Stunting, wasting and underweight in sub-Saharan Africa: A systematic review. Int. J. Environ. Res. Public Health 2017, 14, 863. [Google Scholar] [CrossRef] [PubMed]
  36. Phalkey, R.K.; Aranda-Jan, C.; Marx, S.; Höfle, B.; Sauerborn, R. Systematic review of current efforts to quantify the impacts of climate change on undernutrition. Proc. Natl. Acad. Sci. USA 2015, 112, E4522–E4529. [Google Scholar] [CrossRef] [PubMed]
  37. Akinyele, I.O. Ensuring Food and Nutrition Security in Rural Nigeria: An Assessment of the Challenges, Information Needs, and Analytical Capacity; International Food Policy Research Institute: Abuja, Nigeria, 2009. [Google Scholar]
  38. Brown, M.E.; Pinzon, J.E.; Prince, S.D. Using satellite remote sensing data in a spatially explicit price model: Vegetation dynamics and millet prices. Land Econ. 2008, 84, 340–357. [Google Scholar] [CrossRef]
  39. Grace, K.; Davenport, F.; Funk, C.; Lerner, A.M. Child malnutrition and climate in Sub-Saharan Africa: An analysis of recent trends in Kenya. Appl. Geogr. 2012, 35, 405–413. [Google Scholar] [CrossRef]
  40. Stamoulis, K.; Zezza, A. A Conceptual Framework for National Agricultural, Rural Development, and Food Security Strategies and Policies; Agricultural and Development Economics Division, Food and Agriculture Organization of the United Nations: Rome, Italy, 2003.
  41. Von Braun, J.; Teklu, T.; Webb, P. Famine in Africa: Causes, Responses, and Prevention; The International Food Policy Research Institute: Baltimore, MD, USA; London, UK, 1999. [Google Scholar]
  42. Johnson, K.; Brown, M.E. Environmental risk factors and child nutritional status and survival in a context of climate variability and change. Appl. Geogr. 2014, 54, 209–221. [Google Scholar] [CrossRef]
  43. Chopra, M. Risk factors for undernutrition of young children in a rural area of South Africa. Public Health Nutr. 2003, 6, 645–652. [Google Scholar] [CrossRef] [PubMed]
  44. Griffiths, P.; Madise, N.; Whitworth, A.; Matthews, Z. A tale of two continents: A multilevel comparison of the determinants of child nutritional status from selected African and Indian regions. Health Place 2004, 10, 183–199. [Google Scholar] [CrossRef] [PubMed]
  45. Jolliffe, N. Clinical Nutrition; Harper: New York, NY, USA, 1962. [Google Scholar]
  46. Sastry, N. Family-level clustering of childhood mortality risk in Northeast Brazil. Popul. Stud. 1997, 51, 245–261. [Google Scholar] [CrossRef]
  47. Victora, C.G.; Huttly, S.R.; Fuchs, S.C.; Olinto, M.T.A. The role of conceptual frameworks in epidemiological analysis: A hierarchical approach. Int. J. Epidemiol. 1997, 26, 224–227. [Google Scholar] [CrossRef] [PubMed]
  48. Walker, S.P.; Wachs, T.D.; Grantham-McGregor, S.; Black, M.M.; Nelson, C.A.; Huffman, S.L.; Baker-Henningham, H.; Chang, S.M.; Hamadani, J.D.; Lozoff, B. Inequality in early childhood: Risk and protective factors for early child development. Lancet 2011, 378, 1325–1338. [Google Scholar] [CrossRef] [PubMed]
  49. Dearden, K.A.; Schott, W.; Crookston, B.T.; Humphries, D.L.; Penny, M.E.; Behrman, J.R. Children with access to improved sanitation but not improved water are at lower risk of stunting compared to children without access: A cohort study in Ethiopia, India, Peru, and Vietnam. BMC Public Health 2017, 17, 110. [Google Scholar] [CrossRef] [PubMed]
  50. Rashad, A.S.; Sharaf, M.F. Economic Growth and Child Malnutrition in Egypt: New Evidence from National Demographic and Health Survey. Soc. Indic. Res. 2018, 135, 769–795. [Google Scholar] [CrossRef]
  51. Subramanyam, M.A.; Kawachi, I.; Berkman, L.F.; Subramanian, S.V. Is economic growth associated with reduction in child undernutrition in India? PLoS Med. 2011, 8, e1000424. [Google Scholar] [CrossRef] [PubMed]
  52. Grace, K. Considering climate in studies of fertility and reproductive health in poor countries. Nat. Clim. Chang. 2017, 7, 479–485. [Google Scholar] [CrossRef]
  53. Khatab, K. Childhood malnutrition in Egypt using geoadditive Gaussian and latent variable models. Am. J. Trop. Med. Hyg. 2010, 82, 653–663. [Google Scholar] [CrossRef] [PubMed]
  54. Smith, L.C.; El Obeid, A.E.; Jensen, H.H. The geography and causes of food insecurity in developing countries. Agric. Econ. 2000, 22, 199–215. [Google Scholar] [CrossRef]
  55. Chirwa, E.W.; Ngalawa, H.P. Determinants of child nutrition in Malawi. S. Afr. J. Econ. 2008, 76, 628–640. [Google Scholar] [CrossRef]
  56. Ssewanyana, S.; Kasirye, I. Causes of health inequalities in Uganda: Evidence from the demographic and health surveys. Afr. Dev. Rev. 2012, 24, 327–341. [Google Scholar] [CrossRef]
  57. UNICEF. Nutrition, for Every Child: UNICEF Nutrition Strategy 2020–2030; UNICEF: New York, NY, USA, 2020. [Google Scholar]
  58. Buisman, L.R.; Van de Poel, E.; O’Donnell, O.; van Doorslaer, E.K.A. What explains the fall in child stunting in Sub-Saharan Africa? SSM-Popul. Health 2019, 8, 100384. [Google Scholar] [CrossRef] [PubMed]
  59. Mosley, W.H.; Chen, L.C. An analytical framework for the study of child survival in developing countries. Popul. Dev. Rev. 1984, 10, 25–45. [Google Scholar] [CrossRef]
  60. Habicht, J.-P.; Yarbrough, C.; Martorell, R.; Malina, R.M.; Klein, R.E. Height and weight standards for preschool children: How relevant are ethnic differences in growth potential? Lancet 1974, 303, 611–615. [Google Scholar] [CrossRef] [PubMed]
  61. Niles, M.T.; Emery, B.F.; Wiltshire, S.; Brown, M.E.; Fisher, B.; Ricketts, T.H. Climate impacts associated with reduced diet diversity in children across nineteen countries. Environ. Res. Lett. 2020, 16, 015010. [Google Scholar] [CrossRef]
  62. Kabubo-Mariara, J.; Ndenge, G.K.; Mwabu, D.K. Determinants of children’s nutritional status in Kenya: Evidence from demographic and health surveys. J. Afr. Econ. 2009, 18, 363–387. [Google Scholar] [CrossRef]
  63. Abuya, B.A.; Onsomu, E.O.; Kimani, J.K.; Moore, D. Influence of maternal education on child immunization and stunting in Kenya. Matern. Child Health J. 2011, 15, 1389–1399. [Google Scholar] [CrossRef] [PubMed]
  64. Gewa, C.A.; Yandell, N. Undernutrition among Kenyan children: Contribution of child, maternal and household factors. Public Health Nutr. 2012, 15, 1029–1038. [Google Scholar] [CrossRef]
  65. Ukwuani, F.A.; Suchindran, C.M. Implications of women’s work for child nutritional status in sub-Saharan Africa: A case study of Nigeria. Soc. Sci. Med. 2003, 56, 2109–2121. [Google Scholar] [CrossRef]
  66. Adekanmbi, V.T.; Kayode, G.A.; Uthman, O.A. Individual and contextual factors associated with childhood stunting in Nigeria: A multilevel analysis. Matern. Child Nutr. 2013, 9, 244–259. [Google Scholar] [CrossRef]
  67. Agu, N.; Emechebe, N.; Yusuf, K.; Falope, O.; Kirby, R.S. Predictors of early childhood undernutrition in Nigeria: The role of maternal autonomy. Public Health Nutr. 2019, 22, 2279–2289. [Google Scholar] [CrossRef]
  68. Gayawan, E.; Adebayo, S.B.; Komolafe, A.A.; Akomolafe, A.A. Spatial distribution of malnutrition among children under five in Nigeria: A Bayesian quantile regression approach. Appl. Spat. Anal. Policy 2019, 12, 229–254. [Google Scholar] [CrossRef]
  69. UNICEF. Strategy for Improved Nutrition of Children and Women in Developing Countries; UNICEF: New York, NY, USA, 1990. [Google Scholar]
  70. Rutstein, S.O.; Rojas, G. Guide to DHS Statistics; ORC Macro: Calverton, MD, USA, 2006. [Google Scholar]
  71. Burgert, C.R.; Colston, J.; Roy, T.; Zachary, B. DHS Spatial Analysis Reports No. 7. Geographic Displacement Procedure and Georeferenced Data Release Policy for the Demographic and Health Surveys; Demographic and Health Sureys (DHS) Program: Calverton, MD, USA, 2013. [Google Scholar]
  72. Croft, T.N.; Marshall, A.M.; Allen, C.K. Guide to DHS Statistics; Demographic and Health Sureys (DHS) Program: Rockville, MD, USA, 2018. [Google Scholar]
  73. Leroy, J.L. Zscore06: Stata Command for the Calculation of Anthropometric Z-Scores Using the 2006 WHO Child Growth Standards; Boston College Department of Economics: Chestnut Hill, MA, USA, 2011; Available online: https://ideas.repec.org/c/boc/bocode/s457279.html (accessed on 25 September 2017).
  74. StataCorp Stata Statistical Software; StataCorp LLC: College Station, TX, USA, 2017.
  75. Funk, C.; Peterson, P.; Landsfeld, M.; Pedreros, D.; Verdin, J.; Shukla, S.; Husak, G.; Rowland, J.; Harrison, L.; Hoell, A.; et al. The climate hazards infrared precipitation with stations—A new environmental record for monitoring extremes. Sci. Data 2015, 2, 150066. [Google Scholar] [CrossRef] [PubMed]
  76. Funk, C.; Peterson, P.; Peterson, S.; Shukla, S.; Davenport, F.; Michaelsen, J.; Knapp, K.R.; Landsfeld, M.; Husak, G.; Harrison, L. A high-resolution 1983–2016 T max climate data record based on infrared temperatures and stations by the Climate Hazard Center. J. Clim. 2019, 32, 5639–5658. [Google Scholar] [CrossRef]
  77. Vermote, E.; Justice, C.; Csiszar, I.; Eidenshink, J.; Myneni, R.; Baret, F.; Masuoka, E.; Wolfe, R.; Claverie, M. NOAA Climate Data Record (CDR) of Normalized Difference Vegetation Index (NDVI); NOAA National Centers for Environmental Information: Silver Spring, MD, USA, 2014. [CrossRef]
  78. Paterson, L.; Goldstein, H. New statistical methods for analysing social structures: An introduction to multilevel models. Br. Educ. Res. J. 1991, 17, 387–393. [Google Scholar] [CrossRef]
  79. Roux, A.D. A glossary for multilevel analysis. J. Epidemiol. Community Health 2002, 56, 588–594. [Google Scholar] [CrossRef] [PubMed]
  80. Raudenbush, S.W.; Bryk, A.S. Hierarchical Linear Models: Applications and Data analysis Methods; Sage Publications: Thousand Oaks, CA, USA, 2002. [Google Scholar]
  81. Corrado, L.; Fingleton, B. Multilevel modelling with spatial effects. Economics 2010, 11, 1–21. [Google Scholar]
  82. Goldstein, H. Multilevel models for analysis social data. In Encyclopedia of Social Research Methods; Sage Publications: Newbury Park, CA, USA, 1998. [Google Scholar]
  83. Tobler, W.R. A computer movie simulating urban growth in the Detroit region. Econ. Geogr. 1970, 46, 234–240. [Google Scholar] [CrossRef]
  84. Gelman, A. Multilevel (hierarchical) modeling: What it can and cannot do. Technometrics 2006, 48, 432–435. [Google Scholar] [CrossRef]
  85. ESRI. ArcGIS Desktop; Environmental Systems Research Institute: Redlands, CA, USA, 2017. [Google Scholar]
  86. Olofin, I.; McDonald, C.M.; Ezzati, M.; Flaxman, S.; Black, R.E.; Fawzi, W.W.; Caulfield, L.E.; Danaei, G. Associations of suboptimal growth with all-cause and cause-specific mortality in children under five years: A pooled analysis of ten prospective studies. PLoS ONE 2013, 8, e64636. [Google Scholar] [CrossRef]
  87. Grace, K.; Brown, M.; McNally, A. Examining the link between food prices and food insecurity: A multi-level analysis of maize price and birthweight in Kenya. Food Policy 2014, 46, 56–65. [Google Scholar] [CrossRef]
  88. Shively, G.E. Infrastructure mitigates the sensitivity of child growth to local agriculture and rainfall in Nepal and Uganda. Proc. Natl. Acad. Sci. USA 2017, 114, 903–908. [Google Scholar] [CrossRef] [PubMed]
  89. Sandler, A.M. Against standard deviation as a quality control maxim in anthropometry. Econ. J. Watch 2021, 18, 95. [Google Scholar]
  90. DHS, M. Description of the Demographic and Health Surveys Individual Recode Data File: DHS IV; The DHS Program: Washington, DC, USA, 2008. [Google Scholar]
  91. DHS, M. Description of the Demographic and Health Surveys Individual Recode Data File: DHS V; The DHS Program: Washington, DC, USA, 2012. [Google Scholar]
  92. DHS, M. Description of the Demographic and Health Surveys Individual Recode Data File: DHS VI; The DHS Program: Washington, DC, USA, 2013. [Google Scholar]
  93. Hausman, J.A.; Abrevaya, J.; Scott-Morton, F.M. Misclassification of the dependent variable in a discrete-response setting. J. Econom. 1998, 87, 239–269. [Google Scholar] [CrossRef]
  94. Sandler, A.M.; Rashford, B.S. Misclassification error in satellite imagery data: Implications for empirical land-use models. Land Use Policy 2018, 75, 530–537. [Google Scholar] [CrossRef]
  95. Ghosh, S.; Shivakumar, N.; Bandyopadhyay, S.; Sachdev, H.S.; Kurpad, A.V.; Thomas, T. An uncertainty estimate of the prevalence of stunting in national surveys: The need for better precision. BMC Public Health 2020, 20, 1634. [Google Scholar] [CrossRef] [PubMed]
  96. Rabe-Hesketh, S.; Skrondal, A. Multilevel and Longitudinal Modeling Using Stata; STATA Press: College Station, TX, USA, 2008. [Google Scholar]
  97. Smith, G.D.; Ebrahim, S. Data dredging, bias, or confounding: They can all get you into the BMJ and the Friday papers. BMJ 2002, 325, 1437–1438. [Google Scholar] [CrossRef] [PubMed]
  98. van Wesenbeeck, C.F.A.; Sonneveld, B.G.J.S.; Voortman, R.L. Localization and characterization of populations vulnerable to climate change: Two case studies in Sub-Saharan Africa. Appl. Geogr. 2016, 66, 81–91. [Google Scholar] [CrossRef]
  99. Steyerberg, E.W. Clinical Prediction Models: A Practical Approach to Development, Validation, and Updating; Springer: New York, NY, USA, 2009. [Google Scholar]
  100. Kinyoki, D.K.; Berkley, J.A.; Moloney, G.M.; Odundo, E.O.; Kandala, N.-B.; Noor, A.M. Environmental predictors of stunting among children under-five in Somalia: Cross-sectional studies from 2007 to 2010. BMC Public Health 2016, 16, 654. [Google Scholar] [CrossRef] [PubMed]
  101. Kinyoki, D.K.; Berkley, J.A.; Moloney, G.M.; Odundo, E.O.; Kandala, N.-B.; Noor, A.M. Space–time mapping of wasting among children under the age of five years in Somalia from 2007 to 2010. Spat. Spatio-Temporal Epidemiol. 2016, 16, 77–87. [Google Scholar] [CrossRef] [PubMed]
Nutrients 16 02014 g001
Figure 1. Empirical Bayesian kriging of sample malnutrition prevalence across Kenya and Nigeria DHS-IV, DHS-V, and DHS-VI using ArcGIS Desktop 10.5.1 software by ESRI. Color gradients indicate prevalence of stunting and wasting malnutrition rates.
Figure 1. Empirical Bayesian kriging of sample malnutrition prevalence across Kenya and Nigeria DHS-IV, DHS-V, and DHS-VI using ArcGIS Desktop 10.5.1 software by ESRI. Color gradients indicate prevalence of stunting and wasting malnutrition rates.
Nutrients 16 02014 g001
Nutrients 16 02014 g002
Figure 2. Average marginal effects of categorical determinants of malnutrition (based on Table 5 and Table 6). Variables are displayed such that negative values are beneficial for children’s health and positive values are deleterious for children’s health. The vertical red line at zero marks the liminal threshold, whereas the green and orange horizontal lines are 95% confidence intervals.
Figure 2. Average marginal effects of categorical determinants of malnutrition (based on Table 5 and Table 6). Variables are displayed such that negative values are beneficial for children’s health and positive values are deleterious for children’s health. The vertical red line at zero marks the liminal threshold, whereas the green and orange horizontal lines are 95% confidence intervals.
Nutrients 16 02014 g002
Nutrients 16 02014 g003
Figure 3. Effect of precipitation on average predicted probability of malnutrition. The horizontal axis is the in-sample range of average total monthly rainfall (dm) during the preceding growing season. The horizontal red line demarks the observed malnutrition prevalence and the sloped blue line illustrates how much the expected prevalence rates change as precipitation changes. The shaded blue corresponds to a 95% confidence interval band on the estimate.
Figure 3. Effect of precipitation on average predicted probability of malnutrition. The horizontal axis is the in-sample range of average total monthly rainfall (dm) during the preceding growing season. The horizontal red line demarks the observed malnutrition prevalence and the sloped blue line illustrates how much the expected prevalence rates change as precipitation changes. The shaded blue corresponds to a 95% confidence interval band on the estimate.
Nutrients 16 02014 g003
Nutrients 16 02014 g004
Figure 4. Effect of temperature on average predicted probability of malnutrition. The horizontal axis is the in-sample range of average maximum monthly temperatures (°C) during the preceding growing season. The horizontal red line demarks the observed malnutrition prevalence and the sloped blue line illustrates how much the expected prevalence rates change as temperature changes. The shaded blue corresponds to a 95% confidence interval band on the estimate.
Figure 4. Effect of temperature on average predicted probability of malnutrition. The horizontal axis is the in-sample range of average maximum monthly temperatures (°C) during the preceding growing season. The horizontal red line demarks the observed malnutrition prevalence and the sloped blue line illustrates how much the expected prevalence rates change as temperature changes. The shaded blue corresponds to a 95% confidence interval band on the estimate.
Nutrients 16 02014 g004
Nutrients 16 02014 g005aNutrients 16 02014 g005b
Figure 5. Effect of NDVI on average predicted probability of malnutrition. The horizontal axis is the in-sample range of the unitless NDVI for the three greenest months during the preceding growing season. The horizontal red line demarks the observed malnutrition prevalence and the sloped blue line illustrates how much the expected prevalence rates change as NDVI changes. The shaded blue corresponds to a 95% confidence interval band on the estimate.
Figure 5. Effect of NDVI on average predicted probability of malnutrition. The horizontal axis is the in-sample range of the unitless NDVI for the three greenest months during the preceding growing season. The horizontal red line demarks the observed malnutrition prevalence and the sloped blue line illustrates how much the expected prevalence rates change as NDVI changes. The shaded blue corresponds to a 95% confidence interval band on the estimate.
Nutrients 16 02014 g005aNutrients 16 02014 g005b
Table 1. Summary statistics of discrete variables.
Table 1. Summary statistics of discrete variables.
NigeriaKenya
VariableFrequencyPercentFrequencyPercent
Wasting Status
 Not wasted40,71684.6926,64693.75
 Wasted736015.3117756.25
Stunting Status
 Not stunted29,35361.0620,02570.46
 Stunted18,72338.94839629.54
Sex
 Male 23,99149.9014,36950.56
 Female24,08550.1014,05249.44
Delivery
 Home 29,85062.3814,06949.63
 Clinic 18,00237.6214,27750.37
Birth
 Multiple 14282.977342.58
 Singleton46,64897.0327,68797.42
Weaned
 Breastfed beyond 1 year16,80934.96715825.19
 Weaned by 1 year19,64540.8614,89652.41
 Breastfed up to 1 year11,03822.96417014.67
 Weaned before 1 year5841.2121977.73
Vaccines—Minimum
 No 12,18125.3613414.72
 Yes35,85074.6427,07395.28
Vaccines—Maximum
 No 40,68484.7016,96559.71
 Yes734715.3011,44940.29
Diet
 Unvaried 35,62274.1022,72379.95
 Diverse12,45425.90569820.05
Sick
 Symptomatic12,70926.6614,22650.14
 Asymptomatic 34,95773.3414,14949.86
Latrine—Improved
 No 32,96770.9622,18482.36
 Yes13,48929.04475117.64
Water—Improved
 No 22,08247.0711,54041.37
 Yes24,83352.9316,35558.63
Residence
 Urban15,68032.62817928.78
 Rural32,39667.3820,24271.22
Mother’s Education
 None21,91945.59599221.08
 Primary10,89822.6715,52154.61
 Secondary12,47125.94528018.58
 Higher27885.8016285.73
Wealth Index
 Poorest10,69722.25907731.94
 Poorer10,81322.49578420.35
 Middle967820.13485617.09
 Richer903518.79433315.25
 Richest785316.33437115.38
Interview Month
 January00.0015305.38
 February13702.8512654.45
 March731515.22250.09
 April816616.997292.57
 May870918.12404214.22
 June39328.18471816.60
 July632713.16482816.99
 August569811.85403514.20
 September 40438.41416314.65
 October24855.178052.83
 November310.0611454.03
 December00.0011364.00
Survey Phase
 DHS-IV43869.12471816.60
 DHS-V19,24640.02510117.95
 DHS-VI24,45450.8518,60265.45
Table 2. Summary statistics of continuous variables.
Table 2. Summary statistics of continuous variables.
NigeriaKenya
Standard Standard
VariableAverageDeviationMinMaxAverageDeviationMinMax
Child’s Age (Months)28.317.205928.917059
Mother’s Age (Years)29.56.93154928.66.571549
Birth Tally4.32.581183.82.36116
Precipitation (dm)21.37.954.761.68.36.130.0225.2
Temperature (°C)312.232438.326.43.715.635.6
Precipitation Anomaly0.22.62−11.311.4−0.51.47−5.58.2
Temperature Anomaly−0.70.46−1.90.7−0.80.45−2.60.9
NDVI0.60.140.090.90.60.140.10.9
NDVI Anomaly0.00.026−0.10.20.00.034−0.10.2
Table 3. Hierarchical decomposition of DHS.
Table 3. Hierarchical decomposition of DHS.
NigeriaKenya
Observations per Group Observations per Group
ScaleGroupsMinAverageMaxGroupsMinAverageMax
State 377651299.1275047339600.91165
Cluster 2131122.6792365111.943
Household 30,90411.6820,04811.46
Child48,068 28,241
Table 4. Unconditional hierarchical model—variance decomposition.
Table 4. Unconditional hierarchical model—variance decomposition.
WastedStunted
Hierarchical Fully UnconditionalNigeriaKenyaNigeriaKenya
Variance Decomposition—Percent by Level
States7.09%11.35%10.94%1.87%
Clusters9.48%6.35%6.99%6.11%
Households17.50%20.09%13.31%20.08%
Children65.93%62.22%68.77%71.94%
Table 5. Interpreted hierarchical analyses, wasted percentage point change.
Table 5. Interpreted hierarchical analyses, wasted percentage point change.
Interpreted ResultsPercent Change in Wasted Probability
Hierarchical Random InterceptNigeriaKenya
For a Change from Baseline Category with 95% Confidence Interval in Brackets
Sex—Female−1.2%[−1.9, −0.49]−0.75%[−1.1, −0.36]
Delivery—Clinic−0.91%[−1.7, −0.11]−1%[−1.6, −0.46]
Birth—Singleton−4.1%[−6.7, −1.4]−3.2%[−5.5, −1]
Weaned—By 1 Year Old−0.44%[−1.2, 0.34]−0.11%[−0.48, 0.26]
Vaccines—Minimum−1%[−2, −0.03]−0.44%[−1.4, 0.52]
Vaccines—Maximum−1%[−2, 0]−0.27%[−0.85, 0.32]
Diet—Diverse0.77%[−0.1, 1.6]−0.32%[−0.9, 0.25]
Sick—Asymptomatic−1%[−1.8, −0.25]−0.16%[−0.58, 0.26]
Latrine—Improved−0.31%[−1, 0.38]0.45%[−0.38, 1.3]
Water—Improved−0.26%[−1.2, 0.66]−0.02%[−0.41, 0.37]
Residence—Rural−0.86%[−2.2, 0.47]−0.03%[−0.51, 0.46]
Mothers Education
 Primary−0.96%[−1.8, −0.12]−1.1%[−1.7, −0.56]
 Secondary−2%[−2.8, −1.1]−0.89%[−1.6, −0.18]
 Higher−4%[−5.4, −2.7]−1.7%[−2.6, −0.89]
Wealth Index
 Poorer−0.06%[−0.92, 0.8]−0.92%[−1.6, −0.22]
 Middle−1.3%[−2.2, −0.45]−0.79%[−1.5, −0.04]
 Richer−1.6%[−2.8, −0.42]−1.1%[−1.8, −0.3]
 Richest−0.95%[−2.5, 0.63]−1.2%[−2.3, −0.17]
For a 1-Unit Increase in Determinant with 95% Confidence Interval in Brackets
Child’s Age−2.2%[−2.8, −1.5]−0.13%[−0.38, 0.12]
Mother’s Age0.26%[−0.64, 1.2]−0.22%[−0.65, 0.2]
Birth Tally−0.17%[−0.39, 0.05]0.07%[−0.07, 0.21]
Precipitation−0.96%[−2.3, 0.41]−1.5%[−2.5, −0.63]
Temperature1.2%[0.79, 1.5]0.24%[0.12, 0.36]
Precipitation Anomaly−0.45%[−4.9, 4]1.1%[−0.88, 3.1]
Temperature Anomaly−2.7%[−5.2, −0.26]−0.01%[−0.62, 0.61]
NDVI−9.2%[−14, −4.9]−3.9%[−6.6, −1.3]
NDVI Anomaly4.4%[−14, 23]5.5%[−2.1, 13]
For a 1-Standard Deviation Increase in Determinant with 95% Confidence Interval in Brackets
Child’s Age−3.15%[−4.01, −2.15]−0.18%[−0.54, 0.17]
Mother’s Age0.18%[−0.44, 0.83]−0.14%[−0.43, 0.13]
Birth Tally−0.44%[−1.01, 0.14]0.16%[−0.17, 0.5]
Precipitation−0.76%[−1.83, 0.33]−0.92%[−1.53, −0.39]
Temperature2.68%[1.76, 3.35]0.89%[0.44, 1.33]
Precipitation Anomaly−0.12%[−1.28, 1.05]0.16%[−0.13, 0.46]
Temperature Anomaly−1.24%[−2.39, −0.12]0%[−0.28, 0.27]
NDVI−1.29%[−1.96, −0.69]−0.55%[−0.92, −0.18]
NDVI Anomaly0.11%[−0.36, 0.6]0.19%[−0.07, 0.44]
For a Sample Maximum Increase in Determinant with 95% Confidence Interval in Brackets
Child’s Age−10.82%[−13.77, −7.38]−0.64%[−1.87, 0.59]
Mother’s Age0.88%[−2.18, 4.08]−0.75%[−2.21, 0.68]
Birth Tally−2.89%[−6.63, 0.92]1.04%[−1.08, 3.15]
Precipitation−5.46%[−13.09, 2.33]−3.78%[−6.3, −1.59]
Temperature17.16%[11.3, 21.45]4.8%[2.4, 7.2]
Precipitation Anomaly−1.02%[−11.12, 9.08]1.51%[−1.21, 4.25]
Temperature Anomaly−7.02%[−13.52, −0.68]−0.02%[−2.17, 2.14]
NDVI−7.45%[−11.34, −3.97]−3.12%[−5.28, −1.04]
NDVI Anomaly1.32%[−4.2, 6.9]1.65%[−0.63, 3.9]
Table 6. Interpreted Hierarchical Analyses, Stunted Percentage Point Change.
Table 6. Interpreted Hierarchical Analyses, Stunted Percentage Point Change.
Interpreted ResultsPercent Change in Stunted Probability
Hierarchical Random InterceptNigeriaKenya
For a Change from Baseline Category with 95% Confidence Interval in Brackets
Sex—Female−5.1%[−5.9, −4.2]−7.7%[−9.1, −6.3]
Delivery—Clinic−2.2%[−3.3, −1.1]−4.6%[−6.4, −2.8]
Birth—Singleton−13%[−17, −9.1]−23%[−28, −18]
Weaned—By 1 Year Old−0.31%[−1.6, 1]−1.1%[−2.8, 0.65]
Vaccines—Minimum−0.56%[−2.8, 1.7]−2.9%[−5.5, −0.2]
Vaccines—Maximum−4%[−6, −1.9]−1.6%[−3.1, −0.22]
Diet—Diverse−2%[−3.6, −0.37]−0.51%[−2.3, 1.3]
Sick—Asymptomatic−3.4%[−5, −1.8]−1.3%[−2.6, −0.07]
Latrine—Improved−0.43%[−2, 1.1]−5%[−7.2, −2.8]
Water—Improved0.2%[−1.1, 1.5]−1.1%[−2.7, 0.51]
Residence—Rural1.5%[0.01, 2.9]−1.4%[−3.6, 0.8]
Mothers Education
 Primary−1.5%[−3, −0.01]2.5%[−0.5, 5.6]
 Secondary−5.4%[−7.4, −3.4]−2.5%[−5.7, 0.84]
 Higher−13%[−16, −10]−5.9%[−10, −1.3]
Wealth Index
 Poorer−2.9%[−4.7, −1.1]−4.3%[−6.7, −1.9]
 Middle−6%[−8.2, −3.9]−8.1%[−11, −5.4]
 Richer−12%[−15, −9.9]−10%[−13, −6.9]
 Richest−16%[−18, −13]−16%[−19, −12]
For a 1-Unit Increase in Determinant with 95% Confidence Interval in Brackets
Child’s Age−0.75%[−1.7, 0.16]−2.6%[−3.3, −1.8]
Mother’s Age−3.6%[−4.7, −2.5]−4.6%[−6.1, −3.1]
Birth Tally0.37%[0.06, 0.68]1.1%[0.67, 1.6]
Precipitation−1.5%[−4.4, 1.4]3.3%[0.33, 6.3]
Temperature−0.26%[−1.3, 0.73]−0.92%[−1.2, −0.61]
Precipitation Anomaly5.2%[−1, 11]−3.4%[−8, 1.2]
Temperature Anomaly−1.8%[−5.8, 2.1]1%[−0.43, 2.4]
NDVI−6.6%[−19, 6.1]12%[5.7, 18]
NDVI Anomaly30%[−20, 80]−13%[−36, 9.8]
For a 1-Standard Deviation Increase in Determinant with 95% Confidence Interval in Brackets
Child’s Age−1.08%[−2.44, 0.23]−3.68%[−4.68, −2.55]
Mother’s Age−2.49%[−3.26, −1.73]−3.02%[−4.01, −2.04]
Birth Tally0.95%[0.16, 1.75]2.6%[1.58, 3.78]
Precipitation−1.19%[−3.5, 1.11]2.02%[0.2, 3.86]
Temperature−0.58%[−2.9, 1.63]−3.4%[−4.44, −2.26]
Precipitation Anomaly1.36%[−0.26, 2.88]−0.5%[−1.18, 0.18]
Temperature Anomaly−0.83%[−2.67, 0.97]0.45%[−0.19, 1.08]
NDVI−0.92%[−2.66, 0.85]1.68%[0.8, 2.52]
NDVI Anomaly0.78%[−0.52, 2.08]−0.44%[−1.22, 0.33]
For a Sample Maximum Increase in Determinant with 95% Confidence Interval in Brackets
Child’s Age−3.69%[−8.36, 0.79]−12.78%[−16.23, −8.85]
Mother’s Age−12.24%[−15.98, −8.5]−15.64%[−20.74, −10.54]
Birth Tally6.29%[1.05, 11.56]16.5%[10.05, 24]
Precipitation−8.54%[−25.04, 7.97]8.31%[0.83, 15.86]
Temperature−3.72%[−18.59, 10.44]−18.4%[−24, −12.2]
Precipitation Anomaly11.8%[−2.27, 24.97]−4.66%[−10.96, 1.64]
Temperature Anomaly−4.68%[−15.08, 5.46]3.5%[−1.51, 8.4]
NDVI−5.35%[−15.39, 4.94]9.6%[4.56, 14.4]
NDVI Anomaly9%[−6, 24]−3.9%[−10.8, 2.94]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Sandler, A.; Sun, L. The Socio-Environmental Determinants of Childhood Malnutrition: A Spatial and Hierarchical Analysis. Nutrients 2024, 16, 2014. https://doi.org/10.3390/nu16132014

AMA Style

Sandler A, Sun L. The Socio-Environmental Determinants of Childhood Malnutrition: A Spatial and Hierarchical Analysis. Nutrients. 2024; 16(13):2014. https://doi.org/10.3390/nu16132014

Chicago/Turabian Style

Sandler, Austin, and Laixiang Sun. 2024. "The Socio-Environmental Determinants of Childhood Malnutrition: A Spatial and Hierarchical Analysis" Nutrients 16, no. 13: 2014. https://doi.org/10.3390/nu16132014

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop