Mind the gap: What explains the poor-non-poor inequalities in severe wasting among under-five children in low- And middle-income countries? Compositional and structural characteristics

A good understanding of the poor-non-poor gap in childhood development of severe wasting (SW) is a must in tackling the age-long critical challenge to health outcomes of vulnerable children in low- and middle-income countries (LMICs). There is a dearth of information about the factors explaining differentials in wealth inequalities in the distribution of SW in LMICs. This study is aimed at quantifying the contributions of demographic, contextual and proximate factors in explaining the poor-non-poor gap in SW in LMICs. We pooled successive secondary data from the Demographic and Health Survey conducted between 2010 and 2018 in LMICs. The final data consist of 532,680 under-five children nested within 55,823 neighbourhoods from 51 LMICs. Our outcome variable is having SW or not among under-five children. Oaxaca-Blinder decomposition was used to decipher poor-non-poor gap in the determinants of SW. Children from poor households ranged from 37.5% in Egypt to 52.1% in Myanmar. The overall prevalence of SW among children from poor households was 5.3% compared with 4.2% among those from non-poor households. Twenty-one countries had statistically significant pro-poor inequality (i.e. SW concentrated among children from poor households) while only three countries showed statistically significant pro-non-poor inequality. There were variations in the important factors responsible for the wealth inequalities across the countries. The major contributors to wealth inequalities in SW include neighbourhood socioeconomic status, media access, as well as maternal age and education. Socio-economic factors created the widest gaps in the inequalities between the children from poor and non-poor households in developing SW. A potential strategy to alleviate the burden of SW is to reduce wealth inequalities among mothers in the low- and middle-income countries through multi-sectoral and country-specific interventions with considerations for the factors identified in this study. The Demographic and Health Surveys (DHS) data collected periodically across the LMICs was used for this study. The DHS are cross-sectional in design and are nationally representative household surveys. We pooled data from the most recent successive DHS conducted between 2010 and 2018 and available as of March 2019 and has under-five children anthropometry data. We included only the 51 countries that met these inclusion criteria. The final data consists of 532,680 under-five children living within 55,823 neighbourhoods in 51 LMICs. In all the countries, DHS used a multi-stage, stratified sampling design with households as the sampling unit [31,32]. The DHS computes sampling weights to account for unequal selection probabilities including non-response whose application makes survey findings to fully represent the target populations. The DHS used similar protocols, standardized questionnaires, similar interviewer training, supervision, and implementation across all countries where the survey held. DHS releases different categories of data focusing on different members of households among wish we used the children recode data for the current study. The data covered the birth history and health experiences of under-five children born to sampled women within five years preceding the survey date. The anthropometry measurements were taken using standard procedures [33,34]. The full details of the sampling methodologies are available at dhsprogram.com. The outcome variable in this study is severe wasting. It is defined as “the presence of muscle wasting in the gluteal region, loss of subcutaneous fat, or prominence of bony structures, particularly over the thorax” [35] and approximated by “a very low weight for height score (WHZ) below -3 z-scores of the median WHO growth standards, by visible severe wasting, or by the presence of nutritional oedema” [12] more so, malnutrition has been recently described as “related to both deficiencies and excesses in nutrition, and then, therefore, it includes wasting, stunting, underweight, micronutrient deficiencies or excesses, overweight, and obesity” [36]. SW was a composite score of children weight and height. We generated z-scores using WHO-approved methodologies [37] and categorized children with z-scores <-3 standard deviation as having SW (Yes = 1), otherwise as No = 0. In this decomposition study, household wealth status computed as a composite score of assets owned by households was used as a proxy for family income as DHS does not collect data on family earnings or expenditures. The methods used in computing DHS wealth index have been described previously [38]. Additional details of the methods and assets used for the computation of the wealth quintiles is available at dhsprogram.com. The DHS data had already generated and categorized household wealth quintile as a variable into 5 categories of 20% each: poorest, poorer, middle, richer and richest. For the decomposition analysis, we re-categorized household wealth quintile into two categories: poor (poorest, poorer) and non-poor (middle, richer and richest). A similar categorization has been used elsewhere [8,9,39,40]. Hence, we define “wealth inequality” as “the unequal distribution of assets”. Keywords including low and middle-income countries, childhood morbidity, undernutrition, malnutrition, severe acute malnutrition, severe wasting, were used to search for factors associated with wealth-based inequality in SW across literature database such as PubMed, Medline, Hinari. The individual- and neighbourhood level factors were identified empirically from the literature [11–23,41] are: The individual-level factors are the sex of the children (male versus female): to determine if the biological differences could explain susceptibility to SW; children age in years (under 1 year and 12–59 months): SW has been reported to differ by children ages; maternal education (none, primary or secondary plus): better education could lead to better access to information and enhance earnings, and reduced risk of SW; maternal age (15 to 24, 25 to 34, 35 to 49): younger mothers may have limited education and earnings and thereby increase risk of SW among their children. Others are marital status (never, currently and formerly married): currently married may have spousal support that may reduce the risk of SW; occupation (currently employed or not): capability of providing necessary nutritional intakes; access to media (at least one of radio, television or newspaper): access to information could enhance prevention of SW; sources of drinking water (improved or unimproved), toilet type (improved or unimproved), weight at birth (average+, small and very small), birth interval (firstborn, <36 months and >36 months): children with short birth interval are at higher risk of SW and may have higher experience of wealth-related inequality in SW; and birth order (1, 2, 3 and 4+), children with high birth order are at higher risk of SW and experience higher wealth-related inequality in SW [11–23,41]. We used the word “neighbourhood” to describe the clustering of the children within the same geographical environment. Neighbourhoods were based on sharing a common primary sample unit (PSU) within the DHS data [31,32]. Operationally, we defined “neighbourhood” as clusters and “neighbours” as members of the same cluster. The PSUs were identified using the most recent census in each country where DHS was conducted. We considered neighbourhood socioeconomic disadvantage as a neighbourhood-level variable in this study. Neighbourhood socioeconomic disadvantage was operationalized with a principal component comprised of the proportion of respondents without education (poor), unemployed, living in rural areas, and living below the poverty level [11–23,41]. In this study, we carried out descriptive statistics and analytical analyses comprising of bivariable analysis and Blinder-Oaxaca decomposition techniques using binary logistic regressions. Descriptive statistics was used to show the distribution of respondents by country and key variables. Estimates were expressed as percentages alongside 95% confidence intervals. We computed the risk difference in the development of SW between under-five children from poor and non-poor households. A risk difference (RD) greater than 0 suggests that SW are prevalent among children from poor households (pro-poor inequality). A negative RD indicates that SW is prevalent among children from non-poor households (pro-non-poor inequality). We estimated the fixed effects as the weighted risk differences for each of the country and the random effect as the overall risk difference irrespective of a child’s country of residence. Lastly, the logistic regression method was applied to the pooled cross-sectional data from the 51 LMICs to carry out a Blinder-Oaxaca decomposition analysis (BODA). The BODA is an approach to examine differences in outcomes between groups is the decomposition technique developed by Oaxaca and Blinder [42,43]. This method aims to explain how much of the difference in mean outcomes across two groups is due to group differences in the levels of the independent variables, and how much the difference can be attributed to the differences in the magnitude of regression coefficients [42,43]. The method decomposes the differences in an outcome variable between 2 groups into 2 components so that the gaps between the two groups can be more visible. The first component of the decomposition is the “explained” portion of the gap that captures differences in the distributions of the measurable characteristics (also known as the “compositional” or “endowments”) of these groups. The endowment effect captures differences in the outcome of interest that arises from observed differentials in the characteristics between the groups. Also, the second components of the analysis called the structural or coefficient or return effect, is unexplained and is attributed to differences in the returns to endowments between groups. Thus, each group receives different returns for the same level of endowments. In the analysis of health outcomes, the effect of the return may reflect the indirect effects of structural differences in health systems that affect the healthcare utilization between different groups. In recent time, the classical BODA has been extended from continuous outcomes to binary and other non-linear outcomes [40–43]. We, therefore, adopted this technique to enable the quantification of how much of the gap between the “advantaged” (non-poor) and the “disadvantaged” (poor) groups is attributable to differences in specific measurable characteristics. The non-linear decomposition model assumes that the conditional expectation of the probability of a child having SW is a non-linear function of a vector of characteristics. Using the generalized structure of the model, we fitted a model each for children born to poor and non-poor mothers. The BODA is a statistical method that decomposes the gap in the mean outcomes across two groups into a portion that is due to differences in group characteristics and a portion that cannot be explained by such differences. Therefore, Let A and B be two group names for children from households in poor and non-poor wealth quintiles. Also, let ȲA and ȲB be the mean outcomes for the observations Y in the groups so that the mean outcome difference (ↁȲ) to be explained is the difference between ȲA and ȲB. Then the mean outcome for group G can be written as: where X is a vector containing the predictors and a constant, β contains the slope parameters and the intercept, and ϵ is the error, the mean outcome difference can be expressed as the difference in the linear prediction at the group-specific means of the regressors. That is: Since assuming that E(βℓ) = βℓ and E(ϵℓ = 0). Then the contribution of group differences in predictors to the overall outcome difference can be identified by rearranging Eq 2 to give: In Eq (3), we have divided the outcome difference into three parts thus ↁȲ=E+C+I in the viewpoint of group B so that the group differences in the predictors are weighted by the coefficients of group B to determine the endowment effects. Where E = E(XA) − E(XB)}’βB; is the part of the differentials due to group differences in the predictors that is the “endowment effect”, C = E(XB)′(βA − βB), is the measure of the contribution of differences in the coefficients which includes the differences in the intercept and lastly, I = {E(XA) − E(XB)}′(βA − βB)is the measure of the interaction term accounting for the fact that differences in endowments and coefficients exist simultaneously between the two groups. The E component measures the expected change in group B’s mean outcome if group B had group A’s predictor levels. Similarly, for the C component (the “coefficients effect”), the differences in coefficients are weighted by group B’s predictor levels. That is, the C component measures the expected change in group B’s mean outcome if group B had group A’s coefficients [42,44,45]. In this study, we adopted an alternative (further) decomposition from the concept that there is a nondiscriminatory coefficient vector that should be used to determine the contribution of the differences in the predictors. We assumed β* to be a nondiscriminatory coefficient vector. The outcome difference can then be written as: which can be thought of as ↁȲ=Q+U wherein Q = E(XA) − E(XB)}′β* is the part of the outcome differential that is explained by group differences in the predictors (the “quantity effect”), and the second component, U = E(XA)'(βA − β*) + E(XB)'(β* − βB)is the “unexplained” part. This part is attributed to discrimination, and also captures all the potential effects of differences in unobserved variables. The unknown nondiscriminatory coefficients vector β* can be estimated thereafter by assuming that β* = βA or β* = βB [42], wherein discrimination is directed against A and none against group B, then _βA can be used as an estimate for β* as: and vice-versa. The numerical details have been reported [44,45]. The DHS stratification and the unequal sampling weights of clusters, as well as household clustering effects, were considered. Hence we weighted the data and set significance to 5%. Data were analysed using R statistical software and STATA 16 (StataCorp, College Station, Texas, United States of America). The results of this study are presented in Tables and Figures. All our estimates were weighted. In Table 1, we present the proportion of children from households in the poor wealth quintiles and the prevalence of SW by countries. Also, we present the prevalence of SW among the children from households in the poor and non-poor wealth quintiles within each country. The distribution of the children by the characteristics studied the prevalence of SW by the levels of the characteristics and result of the test of association between the characteristics and the development of SW. *Significant at 0.05 in Mantel Haenszel test of homogeneity of the odds ratio. This study was based on an analysis of existing survey data with all identifier information removed. The survey was approved by the Ethics Committee of the ICF Macro at Fairfax, Virginia in the USA and by the National Ethics Committees in their respective countries. All study participants gave informed consent before participation and all information was collected confidentially. The full details can found at http://dhsprogram.com.