The study context
Ethiopia has a Federal system with nine autonomous Regional States. With an estimated population of 112 million people, the country is dominantly a young population with rapid annual growth rate (about 2.5% per annum). This makes the country the second most populous country in Africa [17]. Ethiopia is the least urbanized country in the world, and has an agrarian economy, where agriculture accounts for more than 60% of the GDP and employs nearly 85% of the population [18]. Severely affected by poverty, food insecurity and communicable diseases, rural women are highly exposed to undernutrition and micronutrient deficiencies [19]. The country experiences one of the highest incidences of child and maternal nutritional deficiencies which contribute to increased morbidity and mortality [12].
Data sources
This study is based on a nationally representative cross-sectional data from the Ethiopian Demographic and Health Survey/EDHS, conducted in 2016 (CSA and ICF, 2016). The EDHS collected sociodemographic data from 10,641 mothers with under five children. The present study used weighted samples of 9949 mothers who were not pregnant at the time and who were aged 15–49 years. The data file contained a wide range of woman’s socioeconomic and demographic variables including key characteristics of her husband.
Ethics statement
EDHS followed previously approved standard protocols, data collection tools and procedures. Participation in the survey was voluntary [12]. Permission to use the data for the purposes of the present study was granted by ICF (U.S.) and Central Statistics Authority (Ethiopia). The ICF and CSA took informed consent from respondents prior to the administration of the questionnaire. Also, written informed consent was obtained from a parent or guardian for participants under 16 years old.
Measures of outcome and exposure variables
The main outcome variables are maternal undernutrition and overweight and/or obesity. Maternal malnutrition was computed as BMI, the ratio of weight (kg) and the square of height (m). Mothers with BMI < 18.5 were defined as having undernutrition. Those with value of 25–29.9 and > 30 were defined as overweight and obese, respectively [3, 20]. Information on parental education was measured as the reported number of years of maternal/paternal education and then allocated within conventional educational categories (e.g., no education, primary, secondary, and post-secondary education). Given the difficulty in generating data on actual household income, EDHS constructed a wealth index from selected key household assets including household ownership of consumer goods, dwelling materials, sources of drinking water, types of sanitation facilities, and other characteristics that relate to economic status [12]. The wealth index was computed by using the Principal Component Analysis (PCA) [21]. This method assigns a weight or factor score, and then standardizes and sums of the scores for each household. The entire sample was then ranked and divided into successive quintiles from the first quintile (Q1 = the poorest 20% of the household population) to the fifth quintile (Q5 = the richest 20% [12, 21].
Other than the main exposure variables described above, seven individual, /household/ community level characteristics were included in the analysis as confounding variables. These were age (15–24, 25–34 and 34 + years), type of residence (rural vs urban), religion (Orthodox Christian vs others), type of family structure (monogamous vs polygamous), work status (working vs not working), household size (small, large and medium), and household headship (male vs female).
Statistical analysis
Descriptive analysis was used to examine the characteristics of the sample. Bivariate logistic regression was conducted to select the variables with p values < 0.2. Multiple logistic regression analyses were conducted to examine the association between selected predictors and the two maternal nutritional morbidity variables, adjusting for confounders. The logistic regression model is given by the equation.
$${text{p}}/left( {{1} – {text{p}}} right) = {text{exp}}left( {{text{a}} + {text{Bx}} + {text{c}}} right)$$
(1)
where P is the probability that the event y occurs, at p(y = 1); and p/[1−p] is the “odds ratio”. We used a p ≤ 0.05 to ascertain statistical significance [22]. Hosmer–Lemeshow test was used to check the goodness of fit in our final model[23]. We used STATA 13 for data management, and data were weighted for descriptive analysis using DHS recommendation.
Inequalities in maternal malnutrition was estimated using a combination of regression-based absolute and relative measuring tools, namely Slope Index of Inequality (SII), Relative Index of Inequality (RII) and Population Attributable Fractions (PAFs). The SII is an absolute measure of the difference of inequality among socioeconomic groups within of a population of interest. The RII is a relative measure, derived from the SII, and considers the size of the population and the relative disadvantage experienced by different subgroups [24]. The computation of the SII and RII started with computing the prevalence of maternal undernutrition and overweight/obese by socioeconomic subgroups (wealth and parental education). Scores were then assigned based on the midpoint range in the cumulative distribution within the population. The SII were estimated by Weighted Least Square (WLS) regression considering the relative rank in the cumulative distribution of the wealth and parental education [24]. The SII is the linear regression coefficient or slope of the regression line given by:
$${text{Y}}_{{{text{ij}}}} = beta_{{0{text{j}}}} + beta_{{{text{1j}}}} {text{X}}_{{{text{ij}}}} + {text{e}}_{{{text{ij}}}}$$
(2)
where Yij: the mean value of overweight/obesity, Xij: the relative rank of the wealth quantile I, β0j: the slope showing the relationship between a group’s and its relative socioeconomic rank. eij: is the error distribution/unexplained error.
The fact that we used socioeconomic groups (analogous to individual ranking data), the regression error term in these Ordinary Least Square (OLS) model presented above becomes less reliable in terms of fulfilling the heteroscedasticity assumption. The relative index of inequality, RII, is derived from SII and the population mean (µ) of the health outcome, given by:
$${text{RII}} = {text{SII}}/mu = beta /mu$$
(3)
Population Attributable Fractions (PAFs) were used to estimate the inequalities in undernutrition across several risk factors to assess the burden at the population level. We used logistic regression estimates to get adjusted PAFs [24,25,26].
The PAFs are directly obtained from logistic regression which was introduced by Greenland and Drescher[25]. The basic idea behind this approach is to estimate a logistic regression model with all known/available risk factors. Ruckinger et al. [26] provided the following steps for calculating the PAF of the risk factor of interest: a)The risk factor has to be coded dichotomously, and then ‘removed’ from the population by classifying all individuals as unexposed, irrespective of their real status b) predicted probabilities for each individual should be calculated using this modified dataset, given by pp = 1/ 1 + exp(−α + β xi); where α represents the estimate for the intercept of the logistic regression model, β denotes the parameter vector for the covariates included in the model, and xi denoting the observations of the covariates for each individual, however, with the ‘removed’ covariate set to zero for all individuals c) Computing the adjusted number of cases of the disease (i.e. overweight or obesity) that is obtained by summing up all predicted probabilities that would be expected if the risk factor was absent in the population, and d) the PAF is then calculated by subtracting these expected cases from the observed cases and dividing by the observed cases[26].
