Hierarchical disentanglement of contextual from compositional risk factors of diarrhoea among under-five children in low- and middle-income countries

Several studies have documented the burden and risk factors associated with diarrhoea in low and middle-income countries (LMIC). To the best of our knowledge, the contextual and compositional factors associated with diarrhoea across LMIC were poorly operationalized, explored and understood in these studies. We investigated multilevel risk factors associated with diarrhoea among under-five children in LMIC. We analysed diarrhoea-related information of 796,150 under-five children (Level 1) nested within 63,378 neighbourhoods (Level 2) from 57 LMIC (Level 3) using the latest data from cross-sectional and nationally representative Demographic Health Survey conducted between 2010 and 2018. We used multivariable hierarchical Bayesian logistic regression models for data analysis. The overall prevalence of diarrhoea was 14.4% (95% confidence interval 14.2–14.7) ranging from 3.8% in Armenia to 31.4% in Yemen. The odds of diarrhoea was highest among male children, infants, having small birth weights, households in poorer wealth quintiles, children whose mothers had only primary education, and children who had no access to media. Children from neighbourhoods with high illiteracy [adjusted odds ratio (aOR) = 1.07, 95% credible interval (CrI) 1.04–1.10] rates were more likely to have diarrhoea. At the country-level, the odds of diarrhoea nearly doubled (aOR = 1.88, 95% CrI 1.23–2.83) and tripled (aOR = 2.66, 95% CrI 1.65–3.89) among children from countries with middle and lowest human development index respectively. Diarrhoea remains a major health challenge among under-five children in most LMIC. We identified diverse individual-level, community-level and national-level factors associated with the development of diarrhoea among under-five children in these countries and disentangled the associated contextual risk factors from the compositional risk factors. Our findings underscore the need to revitalize existing policies on child and maternal health and implement interventions to prevent diarrhoea at the individual-, community- and societal-levels. The current study showed how the drive to the attainment of SDGs 1, 2, 4, 6 and 10 will enhance the attainment of SDG 3. The cross-sectional and nationally representative Demographic and Health Surveys (DHS) data collected during household surveys across most LMIC were used for this study. We extracted and pooled the latest recoded “children data” from the DHS that collected information on diarrhoea, conducted between 2010 and 2018 and available in the DHS data domain by March 2019. Only 57 LMIC met these criteria and were included in this study. The DHS uses a multi-stage, stratified sampling design with households as the sampling unit33,34. However, due to differences in the administrative levels in different countries, the number of sampling stages differed. Country-specific sampling methodologies are available at dhsprogram.com and in the country-specific reports35–37. Sampling weights were computed and provided alongside the data from each country by DHS and were applied to our analysis. The sampling weights were based on the multi-stage sampling procedures to ensure representation of the general population. All the DHS questionnaires were standardized and implemented across all countries with similar interviewer training, supervision, and implementation protocols. The secondary data used for this study is available on request from the owners of the data at https://www.dhsprogram.com/data/dataset_admin/login_main.cfm. Our dependent variable is diarrhoea. Firstly, women were asked to name all births they had within 5 years before the survey dates. They were then asked if any of the children had at least an episode of diarrhoea within 2 weeks preceding the survey date. The response is binary with children who had diarrhoea coded as “1” and “0” otherwise. We used three categories of explanatory variables. Sex of the children (male versus female), children age (< 12 months (infants) and 12–59 months), household head sex (male or female), mothers’ age (15–24, 25–34, 35–49 years), mothers’ highest education (none, primary, secondary or higher); marital status (never, currently or formerly married), employment status (currently employed or not), access to media (yes or no), sources of drinking water (improved or unimproved), toilet type (improved or unimproved), house building material (improved or unimproved), cooking fuel (clean or unclean), weight at birth (average+, small or very small birth weight), and birth order (1, 2, 3 and 4+). These variables have been linked with diarrhoea in the literature10,17,20,31,32. We used the DHS wealth index as a proxy indicator for socioeconomic status. The methods used in computing the DHS wealth index have been described in the literature38 as depicted in Fig. 1. The hierarchical structure of the source data (Authors’ drawings using World Map39). In this study, the terms “neighbourhood” and “community”, were used to describe clustering of children within the same geographical living environment6,40,41. Neighbours are the children that share the same Primary Sample Unit (PSU) within the DHS data. The PSUs were identified using the most recent census in each country where DHS was carried out40,42. The neighbourhood-level factors included in the current study are the place of residence (rural or urban), neighbourhood poverty-, illiteracy- and unemployment levels as illustrated in Fig. 1. The neighbourhood poverty-, illiteracy- and unemployment levels were computed as the proportion of children from households in the lowest two wealth quintiles, children whose mother has no former education and children whose mother was unemployed respectively within each country as of the survey time. We categorized these neighbourhood factors into two levels (low and high) each using the 50th percentile cut-off to allow for non-linear effects and offer useful results for policy decisions. Similar procedures have been used in previous studies40,42. We retrieved the country-level data from the human index reports published by the United Nations database43,44. The Human Development Index (HDI) was created by the United Nations to emphasize “that people and their capabilities should be the ultimate criteria for assessing the development of a country, not economic growth alone”45. The HDI summarizes the average achievement of countries in three key dimensions of human development: “a long and healthy life, access to knowledge and a decent standard of living”45. We categorized the countries into the lowest, middle and highest HDI as shown in Fig. 1. We also explored other country-level factors such as country’s rural area percentages (a measure of the proportion of a countries population that resides in rural areas), multidimensional poverty index (a measure of acute multidimensional poverty) and intensity of deprivation (a measure of the average percentage of deprivation experienced by people in multidimensional poverty)43–45. These variables were used for the descriptive statistics but were excluded from the regression models as they correlated with HDI. We used descriptive statistics to show the distribution of the children by country and by the dependent and independent variables in percentages. Chi-square test of association was used to determine the significance of the association between the independent variables and diarrhoea (Table ​(Table1).1). For the country-level data, we applied sampling weights (SW) provided by the DHS to adjust for unequal cluster sizes, stratifications and to ensure that our findings adequately represent the target population for each country. However, for the pooled data, we computed and applied country-women weights (CWW) to the analysis to reflect the differences in population sizes of the women in each country. The CWW is the product of SW and country-specific weights (CSW). We computed the CSW as the number of sampled women aged 15–49 years divided by the population of women aged 15–49 years for each country. While the number of sampled women is available in the dataset, we obtained the population of each country from United Nations population prospects46. We checked multicollinearity among the independent variables using the “colin” command in Stata version 16. The command provided the variance inflation factor (VIF). All variables with VIF > 2.5 were removed from the regression analysis as literature has shown concerns about VIF > 2.547. Statistical significance was set to 0.05. All analysis was conducted in Stata version 16. Description of Demographic and Health Surveys data by countries and diarrhoea prevalence among under-five children in LMIC, 2010–2018. The multivariable multilevel logistic regression models were used to identify if an association exists between the individual, community contextual factors and national compositional factors and diarrhoea. Using all the 3-level model for binary response specified above, with children i who had diarrhoea (at level 1), from a neighbourhood j (at level 2), and living in a country k (at level 3) as shown in Fig. 1, we identified, constructed and assessed five models to arrive at a robust model that will help identify risk factors of diarrhoea considering the multi-level structure of the data. The models are based on a hierarchical logistic regression model with mixed outcomes consisting of the fixed and random parts as shown in Eq. (1). The probability that a child i of neighbourhood j from country k had diarrhoea is denoted by πijk. The “logit” is the logistic function computed as logitπijk=logπijk1-πijk, β0 is the intercept, βp is the regression coefficient for the p parameters, Xpijk are the covariates, U0jk is the random components due collectively to all children from neighbourhood j of country k while V0k is the random components due collectively to all children from country k. The mixed model enables detailed exploration of variation in variables between higher-level units (contextual heterogeneity). We developed five distinct models to enable a detailed assessment of different combinations of factors to select the most robust model that could identify the contextual and compositional risk factors of diarrhoea. This was aimed at modelling the compositional factors and contextual factors separately and collectively, with reference to the distinct multi-level structure of the data used for the analysis. The first model was the null model (Model I) to assess the variation due to the neighbourhood and country-specific random effects without any explanatory variable. It decomposed the magnitude of variance that existed between country and neighbourhood levels. The second model (Model II) included only the individual-level variables conditional on the neighbourhood and country-specific random effects. The third model (Model III) included only the neighbourhood level variables conditional on the neighbourhood and country-specific random effects. The fourth model (Model IV) examined the country-level variables conditional on the neighbourhood and country-specific random effects, while the final model (Model V), estimated the odds of individual, neighbourhood and country-level variables conditional on the neighbourhood and country-specific random effects. All the models were executed using the multilevel regression model of the MLwinN software, version 3.03 embedded in Stata version 1548. Parameters were estimated using the Bayesian Markov Chain Monte Carlo (MCMC) procedures49 with the following specifications: distribution: binomial; link: logit, burning: 5000, chain: 50,000 and refresh: 500. We reported the results of the fixed effects (measures of association) as the odds ratios (ORs) with their 95% credible intervals (CrIs). Rather than the usual 95% confidence intervals (95% CI) obtained in the frequentist approaches, the Bayesian statistical inference allowed us to summarize probability distributions for measures of association alongside the 95% CrI. The 95% credible interval is simply interpretable as “the 95% probability that the population parameter takes a value in a particular range”. In addition to the fixed effects, we also measured the likely effects of the factors considered across the three different levels using the Intraclass Correlation (ICC) and median odds ratio (MOR). The ICC is the measure of the similarity among children living in the same neighbourhood and within the same country. The ICC is a measure of clustering of odds of having diarrhoea in the same neighbourhood and the same country. We calculated the ICC using the linear threshold, which is the latent variable method50. Adopting the methods recommended by Larsen et. al. on neighbourhood effects51, we reported the random effects in terms of the odds. The MORs are the measures of the variance of the odds ratio in higher levels (neighbourhood and country levels) and it estimates the probability of having diarrhoea that can be attributed to any of the neighbourhood and country factors. If MOR = 1, there is no neighbourhood or country variance. Conversely, the higher the MOR, the more significant are the contextual effects for understanding the probability of developing diarrhoea. A similar approach has been used in similar settings in the literature52,53. This study was based on an analysis of secondary data with all identifier information removed. The Institutional Review Board (IRB) of Inner City Fund (ICF) International Macro at Fairfax, Virginia in the USA reviewed and approved the MEASURE Demographic and Health Surveys Project Phase III. The 2010–2018 DHS’s are categorized under that approval. The Institutional Review Board (IRB) of Inner City Fund (ICF) International Macro complied with the United States Department of Health and Human Services Services guidelines and requirements for the “Protection of Human Subjects” (45 CFR 46). All protocols were carried out in accordance with relevant guidelines and regulations on confidentiality, benevolence, non-maleficience and informed consent. All study participants gave written informed consent before participation and all information was collected confidentially. DHS Program has remained consistent with confidentiality and informed consent over the years. ICF Macro ensures compliance with the U.S. Department of Health and Human Services regulations for the respect of human subjects. The authors sought and obtained express approval to use the data from ICF Macro with Accession number 140625. No further approval was required for this study. The data owners can be contacted at [email protected] and data can be found at https://www.dhsprogram.com/data/dataset_admin/login_main.cfm. Further documentations on ethical issues relating to the surveys are available at http://dhsprogram.com. No patients were involved in the design or dissemination of this analysis.