Demystifying the factors associated with rural–urban gaps in severe acute malnutrition among under-five children in low- and middle-income countries: a decomposition analysis

What explains the underlying causes of rural–urban differentials in severe acute malnutrition (SAM) among under-five children is poorly exploited, operationalized, studied and understood in low- and middle-income countries (LMIC). We decomposed the rural–urban inequalities in the associated factors of SAM while controlling for individual, household, and neighbourhood factors using datasets from successive demographic and health survey conducted between 2010 and 2018 in 51 LMIC. The data consisted of 532,680 under-five children nested within 55,823 neighbourhoods across the 51 countries. We applied the Blinder–Oaxaca decomposition technique to quantify the contribution of various associated factors to the observed rural–urban disparities in SAM. In all, 69% of the children lived in rural areas, ranging from 16% in Gabon to 81% in Chad. The overall prevalence of SAM among rural children was 4.8% compared with 4.2% among urban children. SAM prevalence in rural areas was highest in Timor-Leste (11.1%) while the highest urban prevalence was in Honduras (8.5%). Nine countries had statistically significant pro-rural (significantly higher odds of SAM in rural areas) inequality while only Tajikistan and Malawi showed statistically significant pro-urban inequality (p < 0.05). Overall, neighbourhood socioeconomic status, wealth index, toilet types and sources of drinking water were the most significant contributors to pro-rural inequalities. Other contributors to the pro-rural inequalities are birth weight, maternal age and maternal education. Pro-urban inequalities were mostly affected by neighbourhood socioeconomic status and wealth index. Having SAM among under-five children was explained by the individual-, household- and neighbourhood-level factors. However, we found variations in the contributions of these factors. The rural–urban dichotomy in the prevalence of SAM was generally significant with higher odds found in the rural areas. Our findings suggest the need for urgent intervention on child nutrition in the rural areas of most LMIC. We used cross-sectional data obtained from Demographic and Health Surveys (DHS) conducted between 2010 and 2018 and available as of March 2019—when data analysis started—and that included modules on child health. We chose 2010 to focus on recent surveys in the last decade to allow for comparability. The DHS data are nationally representative household surveys conducted in most LMIC. The surveys have similar methodologies and questions in different countries where the surveys held. The DHS uses a multi-stage, stratified cluster sampling design with enumeration areas as the Primary Sampling Units (PSU) with households at the last stage of sampling22–24. Due to differences in the political and geographical structures across the countries, there are slight variations in the sampling methodologies across the countries. Country-specific sampling methodologies and reports of findings are available at dhsprogram.com25–27. All eligible women and men within each sampled household were interviewed. Sampling weights were calculated based on the population in each stratum to account for unequal selection probabilities whose application makes survey findings to adequately represent the entire population of each country. This is due to the non-proportional allocation of the sample sizes in the different regions and clusters within the same country and the possible differences in response rates. Sampling weights were required for all analysis of the DHS data to ensure the actual representativeness of the survey results at the national levels as well as the sub-national levels. The DHS questionnaires were standardized and implemented across the LMIC with similar interviewer training, supervision, and implementation protocols. The LMIC were determined using the DHS and the World Bank’s categorizations of countries income. For more details, see www.dhsprogram.com; https://data.worldbank.org/income-level/low-and-middle-income and https://datatopics.worldbank.org/world-development-indicators/stories/the-classification-of-countries-by-income.html. The DHS presents the data from each survey in different formats such as women data, child data, birth data, men data, household data etc. In this study, we used the children recode data which was dedicated to health indices of under-five children born to the sampled women within 5 years preceding the survey dates. The severe acute malnutrition is the dependent variable in this study. We defined SAM as “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”4. It is a composite score of children’ weight and height (weight-for-height). The anthropometry measurements were taken using standard procedures. We generated z-scores by applying the WHO-approved methodologies28 to these measurements and categorized children with z-scores < −3 standard deviation as having SAM. SAM is, therefore, a variable with binary outcomes coded as 0 for not having SAM and 1 for those having SAM. The main determinant variable is the place of residence of the children. The mothers’ place of residence was classified as either rural or urban as of the time of the survey by the DHS using standard procedures with minimal differences in what rural areas were across the countries. For more details, see www.dhsprogram.com. The independent variables used in the study were based on the identified factors associated with malnutrition in the literature2,5,16–18,29,30. We categorized the factors into individual-level and neighbourhood-level factors. Individual-level factors include both children, mothers and households characteristics: the sex of the children (male versus female), children age in years (under 1 year and 12–59 months), maternal age (15–24, 25–34, 35–49), occupation (currently working or not), access to media (at least one of radio, television and newspaper), sources of drinking water (improved or unimproved sources), toilet type (improved or unimproved type), weight at birth (average +, small and very small), household wealth index (poorest, poorer, middle, richer and richest), birth interval (firstborn, < 36 months and > 36 months) and birth order (1, 2, 3 and 4 +), recent episode of diarrhoea (yes/no), how soon a child was put to the breast after birth (immediately, within 1 day and after 1 day), availability of health services (whether distances to the health facility was a problem or not), affordability of health services (able to pay for health services or not). However, due to the non-availability of some variables in some countries, the availability and affordability of health services were dropped in the decomposition analysis. The neighbourhood factors were based on the stratum (enumeration areas or geographical clustering) where the children lived. Neighbourhoods were based on sharing a common PSU (enumeration area) within the DHS sampling frame22,23. 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. We computed the neighbourhood socioeconomic disadvantage composite score using principal component analysis of the proportion of respondents within each neighbourhood who are illiterates, poor, and unemployed. In all, data of 532,680 under-five children nested within 55,823 neighbourhoods from 51 LMIC who participated in the DHSs between 2010 and 2018 were analysed. We carried out analytical analyses comprising of univariable analysis, bivariable analysis and multivariate analysis for Blinder–Oaxaca decomposition techniques with binary multivariable logistic regression model. Univariable and bivariable analysis were used to show the distribution of respondents by their countries, the distribution of SAM and the independent variables. We computed the risk difference (RD) in the prevalence of SAM between rural and urban under-five children. Any RD greater than 0 suggests that SAM are more prevalent among children in rural areas (pro-rural inequality). Conversely, a negative RD indicates that SAM is prevalent among children in urban areas (pro-urban inequality). All the descriptive statistics: distribution of characteristics, prevalence and RD were weighted. We computed the random effect of RD in SAM among rural and urban children (Fig. 1). The random effect shows the overall risk difference among all children irrespective of their countries. In Figs. 2 and ​and3,3, we displayed the distribution of RD by countries using colours blue, orange and red to indicate statistically significant pro-rural inequality, no significant inequality and statistically significant pro-urban inequality respectively. Finally, the binary multivariable logistic regression model using the pooled cross-sectional data of SAM from the 51 LMIC was used to carry out a Blinder–Oaxaca decomposition analysis of rural and urban differentials in SAM. Figures 4 and ​and55 shows the decomposition analysis for pro-rural and pro-urban countries rspectively. Forest plot of the risk difference in the prevalence of SAM between rural and urban children by countries. Risk difference between children born to rural and urban mothers in the prevalence of SAM by countries. Scatter plot of rate of SAM and risk difference between children born to rural and urban mothers in LMIC. Contributions of differences in the distribution ‘compositional effect’ of the determinants of SAM to the total gap between children from rural and urban mothers by the pro-rural inequality countries. Contributions of differences in the distribution of ‘compositional effect’ of the determinants of SAM to the total gap between children from rural and urban areas by the pro-urban inequality countries. The Blinder–Oaxaca decomposition is a statistical analysis methodology with an assumption that children born to rural mothers had the same characteristics as children born to urban women31,32. The method allows for the decomposition of the differences in an outcome variable between 2 groups into 2 components so that the gaps can be seen and understood more clearly. It identifies two sources of outcome differentials between groups19,31,33–36. 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. This method enabled the quantification of how much of the gap between the “advantaged” and the “disadvantaged” groups is attributable to differences in specific measurable characteristics. The second component is the “unexplained” part (also referred to as the structural component or return effect) which captured the gap due to the differences in the regression coefficients and the unmeasured variables between the two groups been compared. This second component is attributed to differences in the returns to endowments between groups. So each group had different returns for the same level of endowments19,31,33–36. It was initially built for continuous outcomes but has been extended to analyse non-linear outcomes including binary outcomes which are the most prevalent forms of outcomes in health outcomes and behaviours. For instance, Asuman et al. extended the technique to decompose differentials between rural and urban children who took or did not take immunisation33. In the current study, the non-linear decomposition model assumes that the conditional expectation of the probability of a child having SAM is a non-linear function of a vector of characteristics. The results of the decomposition analysis were presented in Figs. 4 and ​and5.5. The “explained” (compositional component) and the “unexplained” (structural component) portions of the educational inequalities are depicted by red and blue colours respectively; the lighter the red colour, the lower the percentage contribution of the “explained” portion and the lighter the blue colour, the lower the percentage contribution of the “unexplained” portion. All statistical analyses were carried out using Stata 16 and R statistical software. The study flowchart is summarized thus (1) Pool all the data that meet the inclusion criteria (2) determine the prevalence of SAM by rural and urban areas in each country (3) determine the prevalence of SAM by rural and urban areas by the children demographics (4) find the risk differences (the differences in prevalence between rural areas and urban areas in each country) in SAM and display the risk differences for each country to ensure good understanding by the readers (5) determine the pro-rural countries (countries with significantly higher prevalence in rural areas) and the pro-urban countries (countries with significantly higher prevalence in urban areas) and (6) decompose factors associated with rural–urban inequalities in SAM. Ethical approvals were obtained from the Ethics Committee of the ICF Macro at Fairfax, Virginia in the USA and by the National Ethics Committees in the participating countries. Written and signed informed consent was obtained from each parent and/or legal guardians of the children who participated in the study were told that the interviews have minimal risks and potential benefits. All information was collected anonymously and held confidentially. The full ethical approval details have been reported earlier25,27,37 and can be found at https://dhsprogram.com. This study was based on an analysis of existing survey data with all identifier information removed. The surveys were approved by the Ethics Committee of the ICF Macro at Fairfax, Virginia in the USA and by the National Ethics Committees in the participating countries. The full details can be found at https://dhsprogram.com. All methods for data collection and data analysis were carried out following relevant guidelines and regulations on the protection of participants’ data.