Background: Despite continued efforts to address malnutrition, there is minimal reduction in the prevalence rates of stunting in developing countries, including Ethiopia. The association between nutritional and socioeconomic factors collected from a national survey in Ethiopia and stunting have not been rigorously analyzed. Therefore, this study aims to model the effect of nutritional and socioeconomic predictors using 2016 Ethiopian Demographic Health Survey (EDHS) data. Methods: This study is a secondary data analysis of the 2016 EDHS survey, which included 7909 children aged 6 to59 months. Descriptive statistics using frequency and percentage for categorical data and mean and standard deviation for metric data were conducted. Linearity, confounding, and multicollinearity were checked. Bivariable and multivariable logistic regression were carried out. The adjusted odds ratio (AOR) and 95% confidence interval (CI) were calculated. A receiver operative curve was built to estimate the sensitivity and specificity of the model. Results: The study identified that 39.2% of children included in this analysis were stunted. Furthermore, 76.47, 84.27, and 92.62% of the children did not consume fruits and vegetables, legumes and lentils, or meat and its products, respectively. Children aged 24 months to 59 months were found to be at 9.71 times higher risk of being stunted compared to their younger counterparts aged 6-24 months (AOR: 9.71; CI: 8.07, 11.6 children). Those children weighing below 9.1 kg were at 27.86 odds of being stunted compared to those weighing 23.3 kg and above. Moreover, mothers with a height below 150 cm (AOR: 2.01; CI: 1.76, 2.5), living in a rural area (AOR: 1.3, CI: 1.09, 1.54), and being male (AOR: 1.4; CI: 1.26, 1.56) were factors associated with stunting. The predictive ability of the model was 77%: if a pair of observations with stunted and non-stunted children were taken, the model correctly ranks 77% of such pair of observations. Conclusion: The model indicates that being born male, being from a mother of short stature, living in rural areas, small child size, mother with mild anemia, father having no formal education or primary education only, having low child weight, and being 24-59 months of age increases the likelihood of stunting. On the other hand, being born of an overweight or obese mother decreases the likelihood of stunting.
The data from 2016 EDHS was collected from January 2016 to June 2016 in nine geographic regions and two administrative cities of Ethiopia. The survey collects data on demographic and health indicators of all household members with specific emphasis on maternal and child health issues. The sampling frame is based on the Ethiopian Population and Housing Census conducted by the Central Statistical Agency in 2007. The sampling frame is a complete list of 84,915 Enumeration Areas (EA), with each EA comprised of 181 households. Sampling was stratified and conducted at two levels. Each region was stratified into urban and rural, producing 21 strata. Sample EA were selected independently from each stratum in two stages by using proportional allocation and implicit stratification. Accordingly, in the first stage, 202 in urban areas and 443 in rural areas(a total of 645 EAs) were selected with probability proportional to EA size (based on the 2007 Population and Housing Census) and with independent selection in each sampling stratum. In the second stage of selection, a fixed number of 28 households with an equal probability of systematic selection from the newly created list of households per cluster were selected. Detailed sampling procedures and household selection are found elsewhere [11]. Five questionnaires were utilized in the 2016 EDHS – the Household Questionnaire, the Woman’s Questionnaire, the Man’s Questionnaire, the Biomarker Questionnaire, and the Health Facility Questionnaire. All the five questionnaires are presented as an appendix in EDHS 2016 publication [11]. The questionnaires were developed in conjunction with the Demographic Health Survey (DHS), which are customized to an Ethiopian context. Since the DHS recodes the original data into different databases, in this study, we used the children’s recode with the intention to focus our study on children aged 6 to 59 months. The children’s recode contained data mainly about the households’ sociodemographic and other attributes of the children, mothers/caretakers/primary guardians, and fathers/husbands, as well as nutritional, environmental, and health service-related characteristics. In this study, relevant variables were selected based on previous literature reviews, subject matter knowledge, and the objective of the study, which was determining a best fit model of the nutritional and background factors associated with stunting. For model building, to determine factors associated with stunting, the maximum model was specified by considering a thoughtful causal diagram, reducing the number of predictor variables based on descriptive statistics, conducting correlations analysis to remove highly correlated variables, and creation of indices for select variables and conducting bivariable analysis. In addition, variables with more than 15% missing values were only described and not included in model building, the effect of continuous variables was also examined, and continuous predictors were tested for linearity. Based on the above screening procedure, from the sociodemographic and maternal characteristics education of both husband/partner, respondent/ mother highest level of education, age of mother both in a continuous and grouped form, wealth index, residence, sex of child, time to get drinking water, and maternal body mass index (BMI(kg/m2)), weight(kg), and height(cm) and anemia. Lightweight SECA with a digital screen designed and manufactured under the guidance of UNICEF mother-infant scales were used to measure weight. A Shorr measuring board was used to measure height of children. Children younger than 24 months were measured for height while lying down, and older children were measured while standing. Anemia was measured in terms of hemoglobin level in grams/deciliter. Anemia in children is categorized as follows: Details of data collection procedure on the measurement of blood hemoglobin and anemia categorization for children and women are presented in EDHS 2016 [11]. Child size is the mother’s subjective estimate of baby’s size at the time of birth in the 5 years before the survey. By Percent distribution of births by the size of baby at birth, child size is classified into: very small, smaller than average, average, larger and don’t know/missing. This estimate was obtained because birth weight is unknown for most (86%) newborns in Ethiopia [11]. Education was categorized into no education, primary, secondary, and tertiary. Age of the respondents was initially categorized primarily into seven 5-year groups, but, during analysis, it was re-categorized into four groups (15–19, 20–29, 30–39, 40–49). The wealth index indicates a composite measure of a household’s cumulative living standard. It is calculated using easy-to-collect data on a household’s ownership of selected assets, such as televisions and bicycles, materials used for housing construction; and types of water access and sanitation facilities. Wealth index was classified into five groups: poorest, poorer, middle, richer, and richest [11]. However, for the sake of analysis, wealth was re-categorized into 3 groups: poor, average, and rich. The details of construction are attached as supplementary material (S1). Time to get drinking (potable) water was in a continuous form but categorized into less than 30 min or above or equal to 30 min walk. The details of EDHS variables code, including the transformations we made in this study and other details were explained as supplement file. Regarding the dietary intake of children, the data was based on a 24-h recall (day and night before the interview) by the mother who was asked if she had a child living with her who was born after 2014. If her response was affirmative, the mother was asked if she gave the child certain food group selections. Based on this response, we created five groups according to WHO indicators [29] which included: Fruits and vegetables; Grains, roots and tubers; Legumes and lentils; Dairy products; Meat and its products. Table 1 shows the details of categorization and is attached as a supplement file. Sociodemographic and economic characteristics of households (N = 7909) Data were cleaned and analyzed in SAS™ 9.4. Categorical variables were described using frequency and percentages. For continuous variables, mean and standard deviation (SD) were used. Cross-tabulations between some predictor and outcome variables were conducted to check assumptions. Histograms and quartiles were used to present data. For ordinal variables, we used the Spearman correlation. For continuous variables, we used the Pearson correlation to check the correlation between independent predictors, with r ≥ 0.7 used as the cut-off value for correlation. Multi-collinearity was checked using the variance inflation factor (VIF) with VIF < 2.5 used as a cut-off point. Interaction and confounders were tested. The interaction was checked among pairs of variables that were suspected of having interactions based on prior knowledge and literature. A bivariable logistic regression analysis using an Unadjusted Odds Ratio (UOR) was carried out to select candidate variables with P-values of < 0.25 for multivariable logistic regression model building. In multivariable logistic regression analysis model building, backward elimination was used. Finally, variables with a P-value of 0.05 with 95% confidence interval (CI) and adjusted odds ratio (AOR) were conducted. A receiver operative characteristics curve (ROC) with sensitivity and specificity was also depicted to determine the predictive ability of the model. Model goodness-of-fit was assessed by using the Hosmer and Lemeshow test. Linearity was assessed by comparing the squared variable with un-squared variable significance; if the squared variable was significant, the variable was classified into plausible categories. A model selection for non-nested models was done using Akaike’s Information Criteria (AIC), and a model with smaller values was selected.