Study setting
The study was conducted in the north eastern part of Africa in particular reference to Ethiopia. Contextually, the country is categorized as agrarian (Tigray, Amhara,Oromia, SNNP, Gambella, Benshangul Gumuz), pastoralists (Somali, Afar), and city based population (Addis Ababa. Diredawa, Harari).
Data source and study design
The 2019 Ethiopian mini DHS (EMDHS) data were used for this study. Particularly the second Ethiopian EMDHS and the fifth DHS applied were used. The survey was conducted by the Ethiopian Public Health Institute (EPHI) in collaboration with Central Statistical Agency (CSA) and Federal Ministry of Health (FMoH). The nationally representative 8,855 women with in reproductive age (age 15–49) from 8, 663 households were collected from 21 March 2019 to 28 June 2019. A stratified, two-stage cluster random sampling was used in this study. Enumeration areas (EAs) were the sampling units for the first stage of the 2019 EMDHS sample, where 305 of them were randomly drawn. In the second stage, representative samples of households were chosen.
Study variables
Outcome variable
The excusive breast feeding (EBF) practice is an outcome variable, which is categorized as: 1: a mother did not feed the infant/baby anything else except syrup and medicine apart from milk and 0 otherwise. Mothers who have less than 6 months old children during the data collection period were asked about whether they fed breast milk without anything else in the last 24 h preceding the survey, except for oral rehydration salt, syrups, and others therapeutic purposes. As WHO recommendations, exclusive breast feeding for infants should be practiced for the first 6 months.
Independent variables
Two sets of explanatory variables (individual and community level) were included in this study. The individual level variables were taken from infant and their mothers based on the review of different studies. Therefore, the infant related variables are age of infant (0–1 month, 2–3 months, 4–5 months), sex of child (male, female), birth order 9 1st, 2nd − 3rd, 4th and above), time of breastfeeding initiation ( < = 1h, >1h), birth interval ( < = 24 hours, > 24 hours) whereas the maternal socio-demographic and obstetric & healthcare related characteristics are age of mothers in years (15–19, 20–34, 35–49), sex of head of the household (male, female), marital status (never in union, currently in union, widowed/divorced), maternal education ( no education, primary, secondary, higher), religion (orthodox, protestant, muslim, others), region, household wealth index (poor, middle/rich), household family size ( < = 5, > 5), counseling on breastfeeding during the first 2 days of a birth (no, yes), type of birth( single, multiple), ANC visit (no, 1–3, 4+), PNC visit within 2 months (no, yes), place of delivery (home, health facility), and caesarian delivery ( no, yes).
Moreover the community level variables such as place of residence (urban, rural), contextual region (Agrarian, Pastoralist, City based), community ANC (low, high), community PNC (low, high), community poverty (high, low) and community education (low, high) were also included in this study.
Statistical data analysis
Data analysis was done using Stata V14. The data management, such as recoding, renaming and verification were done. To select the candidate variables for final model variable multilevel mixed effect logistic regression model was employed. The multilevel mixed effect logistic regression model was conducted for those variables and a p-value less than 0.05. In multilevel mixed effect logistic regression model, the adjusted odds ration with 95% confidence interval and p-value less than 0.05 were used to identify the most significant factors influencing EBF among under 6-month infants. The assumptions of the model: the presence of multicollinearity among independent variables using variance inflation factor (VIF), categories of the categorical variables using classification table and the model goodness of fit using Hosmer and Lemeshow were checked. Since the EMDHS data are hierarchical (individual were nested within communities), a two-level mixed effects logistic regression model was fitted to estimate both the individual and community level variables (fixed and random) effects on exclusive breastfeeding practices.
Bi-variable and multivariable analysis were computed. In the bi-variable logistic regression, a p-value of less than 0.25 was used to fit three models (Model I: individual level, Model II: community level, Model 3: both individual and multilevel). A total of four models were considered in the multilevel analysis to determine the model which best fits the data; Model I (Null model) without independent variable was employed to evaluate the cluster level, enumeration area, difference in exclusive breast-feeding practice. Model II adjusted for individual level variables to know the contribution of determinants to explain the exclusive breast-feeding practices. Model III used to evaluate the effect community level factors on breast feeding practice of mothers in Ethiopia. Model IV were included in the model to predict the contribution of both individual and community level factors on breast feeding practices.
On the analysis of the model, if the variables had a p-value < = 0.05 with confidence interval not including the null value (OR = 1) were considered as statistically significant variables with EBF practice.
The measures of variation (random effects) between clusters were reported using Intra Cluster Correlation (ICC), Median Odds Ratio (MOR), and Proportional Change in Variance (PCV). The ICC refers to the ratio of the between-cluster variance to the total variance, and it tells us the effect community characteristics on exclusive breast-feeding practices of infants under six months of age in Ethiopia. The ICC can be defined as \(ICC=\frac{{\sigma }^{2}u}{{\sigma }^{2}u+{\sigma }^{2}e}\) or \(ICC=\frac{{\sigma }^{2}}{{\sigma }^{2}+\frac{{\pi }^{2}}{3}}\), where \({\sigma }^{2}u=\) between group variation, \({\sigma }^{2}e=\)within group variation, and \({\sigma }^{2}\) is the estimated variance of the clusters.
MOR is computed as median value of the odds ratio between the randomly selected two areas, the area at the highest likelihood of breast-feeding practice and the area at the lowest likelihood of breast-feeding practice. MOR\(=exp\surd (2x{\sigma }^{2}+0.6745)\approx \text{e}\text{x}\text{p}\left(0.95\sigma \right)\). The proportional change in variance can be calculated as \(PCV=\frac{VA-VB}{VA}x100\%\), where \(VA=\) variance of initial model and \(VB=\) variance of model with more terms measures the total variation attributed by individual and community level factors in the multilevel model. PCV was computed for each model as compared to the null model to know the power of the determinants in the model explains exclusive breast feeding.
The log-likelihood, Akaike Information Criteria (AIC), and deviance were used to select the best model. The model with the highest value of the log-likelihood and with the lowest values of the deviance and AIC was considered to be the best fit model.