Data sources
This study used the Ethiopia Demographic and Health Survey (EDHS) data for the years 2000 (n= 2559), 2005 (n= 1112), 2011 (n= 3569), and 2016 (n= 3106). The data were collected by the Central Statistical Agency (CSA) and Inner City Fund (ICF) International, with funding from the United States Agency for International Development [29] and the Government of Ethiopia [30-33]. The EDHS used a two-stage stratified cluster sampling technique to select the study participants. In stage one, after each administrative region was stratified into urban and rural strata, Enumeration Areas (EAs) were selected using a probability proportional to EA size. In stage two, a household listing operation was carried out in all of the selected EAs and a fixed number of households from each EA were selected [30-33]. All women aged 15–49 years who were permanent residents or who spend the night in the selected households the night before the survey were included in the surveys [30-35]. A weighted total sample of 10,346 women was used, with high response rates that ranged from 94.6% to 97.8%. Detailed methodological strategies used in the surveys have been described elsewhere [30-33]. The present study focused on urban women because past studies have shown that urbanisation is a contributor to the double of malnutrition [20, 21], and women are more likely to be underweight and/or overweight/obese compared to men [14].
Outcome variables
The main outcome variables were underweight and overweight/obesity, measured based on WHO adult body mass index (BMI) classification [15] and used by the Ethiopia Central Statistical Agency and ICF International [29]. BMI was defined as a woman’s weight in kilograms divided by the square of her height in meters (kg/m2). The EDHS used lightweight SECA mother scale to measure weight and Shorr measuring board to assess height [30-33]. BMI was classified into three groups:
- Underweight: BMI < 18.5 kg/m2
- Normal: BMI ≥ 18.5 kg/m2 and BMI ≤ 24.9 kg/m2
- Overweight/obesity: BMI ≥ 25.0 kg/m2
Study variables
The study broadly categorised the study factors as socioeconomic, demographic, behavioural and community-level factors based on previous studies [36, 37]. The selected study factors are associated with underweight and overweight/obesity in reproductive-aged women in previously published studies from LMICs [22, 23, 38-41].
Socioeconomic factors included women’s highest education, women’s employment status, marital status, and household wealth status. Women’s education was classified as ‘no schooling’, ‘primary education or ‘secondary or higher education’. Women’s employment was classified as ‘no employment’, ‘formal employment’ (i.e., professional, technical, managerial, clerical, and services area workers), or ‘informal employment’ (i.e., agricultural and manual workers) [36, 42]. Marital status was classified as ‘never married’, ‘formerly married’ or ‘currently married’. The EDHS used principal components analysis (PCA) to calculate the household wealth index based on a series of variables relating to ownership of household assets such as television and bicycles; type of materials used for housing construction; and types of water source and sanitation facilities [43]. The household wealth index was classified as ‘poor’, ‘middle’ or ‘rich’, consistent with previously published studies [44, 45].
Demographic and behavioural factors included women’s age, parity, listening to the radio, reading newspapers/magazine, and watching television. Women’s age was classified as ‘15–24 years’, ‘25–34 years’ or ‘35 and above years’, and women’s parity classified as ‘none’, ‘1–4 children’ or ‘5 or more children’. Women who reported exposure to the media (radio, magazine/newspaper or television) at least once a week were classified as ‘Yes’ and those who did not were classified as ‘No’. Community-level factor (i.e. region of residence) was classified as ‘Tigray’, ‘Afar’, ‘Amhara’, ‘Oromia’, ‘Somali’, ‘Benishangul’, ‘Southern Nations Nationalities and Peoples Region (SNNPR)’, ‘Gambella’, or ‘Metropolis’ regions based on Ethiopia’s geopolitical and administrative features, consistent with the EDHS report and previously published studies [30-33, 36]. The Metropolis region included Addis Ababa and Dire Dawa city administrations, and the Harari region. Among the study participants, about 44.1% of women had no employment, and nearly half (47.1%) of them were in the 15–24 years’ age group [Additional file 1].
Statistical analysis
Preliminary analyses involved the description of the study participants by calculating frequencies and percentages of the study variables. This was followed by the estimation of the prevalence of the outcome variables (underweight and overweight/obesity) and by the selected study variables (socioeconomic, demographic, behavioural and community-level factors) in both year-specific data (2000, 2005, 2011 and 2016) and in the combined dataset. We used the combined dataset to increase the statistical power of the study in order to detect any association between the study factors and the outcomes, as well as to examine trends in underweight and overweight/obesity over the study period (2000–2016).
Multivariate multinomial logistic regression modelling was used to examine the association between socioeconomic, demographic, behavioural and community-level factors and (i) underweight and (ii) overweight/obesity using the normal weight group as a reference category. Specifically, socioeconomic factors were entered into the model to assess their relationship with the outcomes, with adjustment for demographic, behavioural and community-level factors (stage 1). A similar strategy was used in models of demographic factors to examine their relationship with the outcome variables, with additional adjustment for socioeconomic, behavioural and community-level factors (stage 2). Similar modelling techniques were used for the behavioural and community-level factors in the third and fourth stages (stages 3 and 4), respectively.
In the models, we adjusted for the survey years in the combined dataset, while sampling weight and clustering were accounted for in both the year-specific and combined datasets. Collinearity was checked using ‘variance inflation factor (VIF)’ but no significant results were evident in the analyses. We also estimated P for trends in each category of the study variables to assess for any convergence or divergence. Odds ratios with 95% confidence intervals (CIs) were estimated as the measure of association between study factors and outcome variables. All statistical analyses were conducted using Stata version 14.0 with `svy’ command to adjust for sampling weights, clustering effects and stratification, and the `mlogit’ function was used for the modelling.