Study setting
This study is a further analysis of nationwide data, Ethiopian Demographic and Health Survey (EDHS) data, which was collected between January 18, 2016 and June 27, 2016 [19]. The 2016 EDHS is the fourth national representative cross-sectional survey to be conducted as part of the global MEASURE DHS initiative, which is led by the Ethiopian Central Statistical Agency (CSA). The target population for this study was pairs of all 6–59 month-old children with their mothers or caregivers in Ethiopia. Furthermore, the study population was pairs of all 6–59 month-old children with their respective mothers or caregivers in the randomly selected enumeration areas (EAs) of Ethiopia.
Sampling procedures
Briefly, the 2016 EDHS employed stratified, two-stage cluster sampling to identify the representative samples. The sampling frame for the 2016 EDHS consists of a total of 84,915 Enumeration Areas (EAs). In the first stage, 645 (202 urban and 443 rural) EAs were chosen. Figure 1 presents the map of survey cluster (EA) locations where raw dataset were collected. Then, in the second stage, a fixed number of 28 households was chosen from each enumeration area. A total of 16,650 households were included in the survey. A nationally representative population of 9,504 children aged 6–59 months in the chosen households and a total of 15,683 women aged 15—49 years were interviewed with a 95% response rate. A thorough description of the survey design and sampling procedure can be found elsewhere [19].
The unit of analysis was the child-mother/caregiver pair. The primary and secondary sampling units were clusters and households, respectively. To account for non-response, post-stratified weights were considered [20, 21]. For selected survey years, sociodemographic and anthropometric data from children under the age of five and their mothers (aged 15–49 years) were retrieved. Hence, in this study, 7,624 under-five child-mother/caregiver pairs with complete anthropometric and hemoglobin records were included.
Variables of the study
In this study, we looked at two binary outcome variables. Specifically, the double burden of malnutrition (DBM) and triple burden of malnutrition (TBM) among child- to- mother/caregiver pairs.
DBM among child – mother/caregiver- pairs was defined based on related literatures [22,23,24,25]. That is, Yi = 1 when a child is undernourished (either stunted, wasted, or under-weight based on the World Health Organization (WHO) child growth standards [26]) and the mother/caregiver is over-nourished (overweight/obese), and Yi = 0 when neither is the case. Using the WHO standards of BMI (weight (kg)/height (m2)) [27], maternal nutritional status was classified as underweight (≤ 18.4); normal (18.5–24.9); overweight (25.0–29.9); and obese (≥ 30 kg/m2).
Based on related studies, the TBM was defined as a combination of the DBM of a child-mother pair plus an anemic child [22]. TBM includes undernourished children (either underweight, stunted, or wasted), an anemic child (with a micronutrient deficit, which is frequently caused by iron deficiency), and an over-nourished mother/caregiver (with a weight higher than healthy for height, either overweight or obese).
The double and triple burden of malnutrition among child-mother pairs in Ethiopia is the result of a number of complicated, multifarious, and interconnected variables that work at several levels, one of which is maternal-related characteristics. Accordingly, the independent variables were selected based on literature and their availability in the data used. Household (HH), mother and child characteristics are among the independent variables considered in this study. Specifically, residence (urban or rural) [22, 25, 28], child age [22, 25, 28], HH wealth [22, 23, 25, 28], mother’s age [22, 28], mother’s education level [22, 23, 28], and child sex [22, 28]. Other socio-demographic and environmental factors include employment status, marital status, family size, source of drinking water, and type of toilet facility [22].
Data processing and analysis
The EDHS 2016 survey data sets and the Global Positioning System (GPS) points were obtained and processed with permission from Measure DHS (http://www.dhsprogram.com). The variables were then retrieved from the survey data for the children’s data sets.
For spatial analysis, ArcGIS Pro version 2.4 was utilized, while R was used for the remaining analyses. Because of the nature of the sampling design, all analyses were performed utilizing the complex sampling design adjustment approach and non-response rate. A survey package of R [29] was used to estimate confidence intervals around prevalence by taking sample weights into account, which represent the inverse of the chance that the observation is included.
To estimate the prevalence of DBM and TBM across socio-demographic determinant variables, descriptive statistics were utilized. The prevalence of overweight/obese mother and stunted child (OM/SC), overweight/obese mother and wasted child (OM/WC), overweight/obese mother and underweight child (OM/UC), overweight/obese mother and anemic child (OM/AC), DBM, and TBM were presented as weighted percentages with 95% confidence intervals (CIs).
Further, the GPS coordinates were then combined with the prevalence of DBM and TBM in each of the EDHS 2016 clusters. As a result, the cluster level prevalence of DBM and TBM was exported into ArcGIS to depict hot and cold spots of clusters. Geographic variation in DBM and TBM prevalence among EDHS clusters was identified using spatial analysis [30, 31]. Geographic variation of significant high prevalence or low prevalence of DBM and TBM was computed for each cluster using the Moran’s I statistic [30]. Maps depicting the distribution and variations of DBM and TBM throughout the country were constructed. In addition, as a complement to Moran’s I statistic, inverse distance weighted interpolation [31] was employed to estimate these distributions.
The standard binary logistic regression estimates are inadequate for a data from a complex survey design since the data originates from a complex survey design with stratification, clustering, and unequal weighting [32]. If the complex survey design is not included in the analysis, the standard errors are likely be underestimated, perhaps leading to statistically significant findings when they are not [20, 21, 32]. As a result, the survey binary logistic regression model [32] was used to analyze data in order to account for the complex sampling design. The bivariate and multivariable survey binary logistic regression models were used to evaluate the associated factors of the DBM and TBM. Bivariate analysis was used to examine the relationship between socio-demographic characteristics and outcome variables. The multivariable analysis included all variables with p-value less than or equal to 0.25 in the bi-variable analysis. Variables with a p-value of < 0.05 were considered statistically significant in multivariable analysis.
To demonstrate the strength of the association, the adjusted odds ratio (AOR) with the accompanying 95% CIs were provided.