Background: Understanding the drivers contributing to the decreasing trend in stunting is paramount to meeting the World Health Assembly’s global target of 40% stunting reduction by 2025. Methods: We pooled data from 50 Demographic and Health Surveys since 2000 in 14 countries to examine the relationships between the stunting trend and potential factors at distal, intermediate, and proximal levels. A multilevel pooled trend analysis was used to estimate the association between the change in potential drivers at a country level and stunting probability for an individual child while adjusting for time trends and child-level covariates. A four-level mixed-effects linear probability regression model was fitted, accounting for the clustering of data by sampling clusters, survey-rounds, and countries. Results: Stunting followed a decreasing trend in all countries at an average annual rate of 1.04 percentage points. Among the distal factors assessed, a decrease in the Gini coefficient, an improvement in women’s decision-making, and an increase in urbanization were significantly associated with a lower probability of stunting within a country. Improvements in households’ access to improved sanitation facilities and drinking water sources, and children’s access to basic vaccinations were the important intermediate service-related drivers, whereas improvements in early initiation of breastfeeding and a decrease in the prevalence of low birthweight were the important proximal drivers. Conclusions: The results reinforce the need for a combination of nutrition-sensitive and-specific interventions to tackle the problem of stunting. The identified drivers help to guide global efforts to further accelerate stunting reduction and monitor progress against chronic childhood undernutrition.
We considered data from DHSs conducted in 42 LMICs countries that are partners for nutrition support by the European Commission’s Directorate-General for International Cooperation and Development. Our trend analysis required data from at least 3 time-points. Therefore, the present analysis only utilized data from 14 countries with an adequate number of standard DHS rounds since 2000. Countries included in our analysis were Bangladesh, Cambodia, Ethiopia, Haiti, Kenya, Malawi, Mali, Nepal, Nigeria, Rwanda, Tanzania, Uganda, Zambia, and Zimbabwe. The DHS has been conducted in several countries for over three decades, providing nationally representative cross-sectional data on demographic, health, and nutrition information, among others [26]. DHS uses standardized data collection procedures across countries and consistent content across rounds, allowing for comparison of data across-countries and within-country over time. We pooled data from 50 DHS rounds in the 14 countries. There were a different number of surveys per country and different time spans between surveys for the countries. Data were compiled from the official website of the DHS Program (https://dhsprogram.com/) accessed with permission in November 2018. Children under five years of age with height measurements and other relevant information available were included in the analysis. The outcome of interest was the probability of being stunted in under-five children. Stunting here is described as binary (stunted/not stunted), defined as a height/length-for-age z score below 2 SD from the median based on the WHO 2006 Child Growth Standards [27]. Potential drivers for stunting reduction were considered based on UNICEF’s conceptual framework for the determinants of childhood undernutrition [28] and the Lancet review for the framework of action [29]. A variable was selected for analysis when it was available in the DHS datasets for enough survey rounds, used as a determinant in more than one paper, or mentioned in reviews and/or were found to be significant in previous country analyses. We considered three groups of variables operating at different levels as distal factors, intermediate health and related services utilization factors, and proximal factors (Table 1) [25]. We considered variables like education coverage and women’s decision-making power and work opportunity as distal factors because these factors can influence access to and utilization of the intermediate service-related variables. Definitions of our potential drivers are in line with the DHS statistics guideline [30]. Additionally, the associations between drivers and stunting were adjusted for important demographic and socioeconomic covariates for the child, mother, and household, including child age, sex, birth-order and birth-interval, maternal age and marital status, household wealth status, and place of residence (urban/rural). Description of variables used in the study 1. 1 BCG, Bacille Calmette Guerin vaccine against tuberculosis; DPT, diphtheria-pertussis-tetanus vaccine; DHS, Demographic and Health Survey; MCV, Measles antigen-containing vaccine; SD, standard deviation; UNICEF, United Nations Children’s Fund; WHO, World Health Organization. Data management and statistical analyses were conducted in Stata version 14.1 (StataCorp LLC, College Station, TX, USA) using the High Performance Computing infrastructure at Ghent University. For all analyses, associations with a p-value of less than 0.05 were regarded as statistically significant. Missing indicators for certain survey-rounds were imputed based on linear interpolation of the available time-points within a country. The weighted prevalence and average values of stunting and explanatory indicators in each survey-round were calculated considering the DHS sampling weight factor using the svyset command. Descriptive evaluation of the trend in stunting and the explanatory indicators over survey-rounds within-country was conducted using locally weighted scatterplot smoothing graphs. The pooled annualized rate of change in stunting prevalence and explanatory variables were estimated by fitting mixed-effects linear regression models with a random intercept country and as a random slope survey round, and applying weightings based on the population size of each country. We followed the approach by Fairbrother [31] to analyze repeated cross-sectional survey datasets from different subjects. The method allows for using individual-subject level data to examine the association between aggregate level data on potential drivers and stunting risk for an individual child, while adjusting for time trends (survey year) and important demographic and socioeconomic covariates for the child, mother, and household mentioned above. A mixed-effects linear probability model with a robust variance estimator was fitted using four-level random intercepts to account for potential sources of clustering in our data where individuals (level-1) are nested within the DHS sampling clusters (level-2), which in turn are nested within survey-rounds (level-3), and finally, survey-rounds nested within countries (level-4). Since an indicator can vary both within-country, over time, and across-country, we sought to identify separate within-country (longitudinal) and between-country (cross-sectional) components for the association between an indicator variable and stunting. For this purpose, for each indicator variable, we calculated the average value per country (representing cross-country differences) and the average value per survey-round in a country mean-centered (representing within-country change over time), which were simultaneously specified in the model. The within-between model specification has the additional advantage of minimizing potential endogeneity problems in mixed-effects models [32]. In order to avoid over-adjustment from the inclusion of distal, intermediate, and proximal indicators in a single model, we specified three separate multivariable regression models for each group of indicators. The relationships between an explanatory variable and stunting in the four-level models are represented using the equation below: where yiktj is the risk of stunting in child i who is from a DHS sampling cluster k nested within a country-year tj, which in turn is nested within a country j. β1 and β2 give estimates for stunting-indicator associations decomposed into between-country (which is estimated from the value of an indicator at a country level using the average of all survey-rounds in a country (x̅j)) and within-country effects (which is estimated from the average value of an indicator for each survey-round per country (xtjM)), respectively. Country-year level variables were mean-centered by subtracting the value of the country-level variable from the value at survey-round within the same country. β3 gives the estimate for the association of individual-subject level covariates used for adjustment in the model. β4 gives the estimate for the time variable as a set of year dummies, which adjust for possible simultaneous but unrelated time trends in both xj/xtjM and stunting. Random effect components modeling clustering of data by the DHS sampling cluster (uk), by country-years (ukj), and countries (uktj), and the residual term for individual child (εiktj) are assumed to follow a Gaussian distribution with a mean value of zero.