Study setting and data source
Ethiopia is the study area. Administratively, Ethiopia is divided into nine regional states: Tigray, Afar, Amhara, Oromia, Somali, Benishangul, SNNPR, Gambella, Harari and two city administrations: Addis Ababa and Diredawa. The data source is the nationally representative 2011 Ethiopia Demographic Health Survey (EDHS). The survey was a population-based cross-sectional study designed to provide population and health indicator estimates at national and regional levels, as well as urban and rural residents.
Sample size and sampling procedures
Data from the EDHS 2011 were used, specifically data on individual women of childbearing age. All eligible women in the 624 clusters were the study population. The sample was selected using a stratified, two-stage cluster design and enumeration areas (EAs) were the sampling units for the initial stage of sampling. The sampling frame was a list of all EAs established from the population and housing census in 2007. The first stage involved the selection of clusters. The second stage involved the selection of households from the selected clusters. Following the above procedures at the first stage, the sample contained 624 EAs, but 28 of the clusters were not interviewed because of the drought and security problems in the Somali region. In the second stage, a representative sample of 17,817 households was selected for the survey with 17,385 eligible women identified for individual interviews, and 16,515 women were interviewed. To gain interpretability of results, those who answered don't know and had a missing response for all justifications were excluded. These exclusions resulting in a loss of only 149 (0.9%) women and giving a final sample of 16,366 for the analysis.
Study variables and measurements
In lower-income countries, including Ethiopia women’s acceptance of IPV were measured using attitudes toward IPV scale of measurement as recommended by the DHS measure [35]. The justification was measured in each survey question by assessing response (yes/no) to five attitudinal scenarios/questions. Women were asked if they felt a husband would be justified in beating his wife if she: goes out without telling him, neglects the children, argues with him, refuses to have sex with him and burns the food. Responses to these questions were transformed into a single dichotomous "Yes" or "No" variable. Women who responded "Yes" to one or several of the questions formed and were coded as Yes (1) and women who responded "no" to all the questions coded as No (0).
The independent variables were socio-economic and demographic characteristics of the respondents (women’s education, literacy, partner education, education difference, partner's occupation, women’s occupation, owning a house, wealth index, ever chewed chat, alcohol consumption, women’s autonomy, marital status, family system, women’s age, age at first sex, age at first cohabitation, partner age, number of living children, cohabitation duration, pregnancy status; cultural factors: ethnicity and religion); psychosocial factors (perceived existence of law); and Community-level factors (community literacy, community poverty, community media, community residence and State region).
Women empowerment is measured by women’s participation in household decision making concerning who decides on: women's health care, large household purchases, visits to family or relatives and how men's earnings are used were measured in the DHS. If the woman decided jointly with her partner or by herself, she was assigned as participated in decision making and did not otherwise. Further, a new variable 'women empowerment' was created by assuming participation as a proxy measure of women empowerment and leveled into: Empowered if she is involved in four of the decision making, Partially empowered if involved in one of the decisions, two of the decision and three of the decision, and not empowered if not involved in any decision.
Community-level variables were created by aggregating individual's characteristics within their clusters. They were computed using the proportion of selected levels of a given variable that were concerned with per cluster. Since the aggregate values for all generated variables have no meaning at the individual-level, they were categorized into groups based on the national median values. Through this aggregation, the proportion of community factors ranging between 0 and 50th percentiles were categorized as low, and the range between 50 and 100th percentiles were categorized as high. Median values were used because of the non-normality of aggregated variables. Community poverty was constructed from the first two lower quintiles (poorest and poor) as proportions, and distinguishing clusters with low (0–50th percentiles) and high level of community poverty (50–100th percentiles). This procedure was also applied to create community-level factors for community media exposure considering the proportions of community members who have been exposed to any media (listening to the radio, watching television, reading magazines or newspapers) and community literacy (proportion of individuals who were able to read the whole sentence among women in the specified cluster). The two non-aggregate community-level factors included: residence (urban and rural), and contextual region dichotomized into city administration and State region.
Statistical analysis
The DHS variable recode was designed to standardize variables that would make cross-country analysis easier and comparable. Distribution and values for each variable were assessed to detect implausible values and missing data values managed accordingly. Data were cleaned and analyzed using STATA software version 12.0. Data were examined and summarized using frequency and percent and presented using a table and bar graph. To get a reliable estimate data was given weight to adjust for differences in the probability of selection and non-response. Bivariate multilevel mixed-effects binary logistic regression was used for analyzing the association between explanatory variables and women’s acceptance of IPV. Variables with a p-value less than 0.05 in the bivariate analyzes were candidates for the multivariate analysis.
Multivariate two-level mixed-effects logistic regression was applied to the data to predict a binary outcome variable from a set of individual and community-level independent variables. The 2011 EDHS data present a clear multilevel structure and multilevel modeling used to permit the inclusion of error terms that reflect the variation pattern introduced by the data's hierarchical structure. Therefore, this analytic method was employed to account for the hierarchical structure of the data, in which 16,366 individuals (level 1) nested within 596 community groups (level 2).
The proportions of total variance related to community level factors were estimated by the intraclass correlation coefficient (ICC). The proportional change in variance (PCV) is the percentage reduction from the estimated variance in the null model as a result of included independent variables in the model. Results of fixed effects were interpreted with an adjusted odds ratio (AOR) with a 95% confidence interval (95%CI). The random effect was interpreted using ICC and PCV and compared across the progressive models by looking at them.
The interaction effect was checked and there was no interaction effect (“Appendix 3”). Moreover, the multicollinearity was also checked by using variance inflation factors (VIF) and no variable had VIF > 10 [36, 37]. Akaike information criterion (AIC) was used to compare models with different sets of parameters. A model with the lowest Akaike Information Criteria (AIC) was considered as the best fit model.
Data quality assurance
Standard model questionnaires were designed and developed by the DHS program with the basic approach of collecting quality data. Developed English version questionnaires were translated into three major languages Amharigna, Afan Oromo, and Tigrigna. Complete interviews were conducted, yielding a response rate of 95%.