Spatial distribution and determinants of intimate partner violence among reproductive-age women in Ethiopia: Spatial and Multilevel analysis

This study was based on the 2016 Ethiopian Demographic and Health Survey (EDHS) data. EDHS is a nationally representative survey conducted in every five years interval in Ethiopia. Ethiopia has nine regional states (Afar, Amhara, Benishangul-Gumuz, Gambela, Harari, Oromia, Somali, Southern Nations, Nationalities, and People’s Region (SNNP) and Tigray) and two Administrative Cities (Addis Ababa and Dire-Dawa). A stratified two-stage cluster sampling technique was employed to select the study participants. At the first stage, a total of 645 Enumeration Areas (EAs) were selected. In the second stage, on average 28 households per EA were selected. Overall, for EDHS 2016 a total of 18,008 households were chosen and 16,583 eligible women in the selected household were identified. For this study, a total weighted sample of 6090 women were included. The detailed sampling procedure has been presented in the full report EDHS 2016 [31].

ArcGIS version 10.6 and SaTScan version 9.6 statistical software were used to explore the spatial distribution and to identify the hotspot areas of intimate partner violence. The spatial global autocorrelation (Global Moran's I) was used to determine whether intimate partner violence was randomly distributed or not [29]. Moran's I is a spatial statistic used to measure autocorrelation in space by taking the entire data set and generating a single output value ranging from -1 to + 1. A statistically significant Moran's I value (p < 0.05) indicates that the spatial distribution of intimate partner violence is non-random and suggests the existence of spatial autocorrelation. Besides, Getis-OrdGi * statistical hotspot analysis was done to identify significant hotspot and cold spot areas of intimate partner violence [32]. The Bernoulli based model spatial scan statistical analysis was conducted to identify significant primary and secondary clusters of intimate partner violence. The SaTScan uses a circular scanning window that goes across the region of the study. Women who had experienced intimate partner violence were considered cases while those who had not experienced intimate partner violence were taken as controls to fit the Bernoulli model. The default overall spatial cluster size of < 50% of the population was used as an upper limit, allowing for the identification of small and large clusters and excluding clusters that contained more than the maximum limit. A likelihood ratio test statistic and the p-value were used to determine significant clusters for each possible cluster. The most likely performing cluster was the scanning window with a maximum likelihood. The primary and secondary clusters were established and ranked based on their likelihood test, based on 999 replicates from Monte Carlo [33].

The Kriging spatial interpolation technique was applied to predict the prevalence of IPV in un-sampled/unmeasured areas based on the values observed from sampled areas. There are various deterministic and geostatistical interpolation methods [34]. For this study, the Ordinary Kriging spatial interpolation method was used since it had a smaller residual and root mean square error.

The data were weighted using sampling weight, primary sampling unit, and strata before any statistical analysis to restore the representativeness of the survey and to take into account the sampling design to get reliable statistical estimates. Descriptive and summary statistics were conducted using STATA version 14 software. The EDHS data has hierarchical nature and women are nested within a cluster and we expect that women within the same cluster may be more similar to each other than women in the rest of the country. This violates the assumption of the traditional regression model which is the independence of observations and equal variance across clusters. This implies that the need to take into account the between cluster variability by using an advanced model. Therefore, a multilevel random intercept logistic regression model was fitted to estimate the association between the individual and community level variables and the likelihood of experiencing intimate partner violence. Model comparison was done based on Deviance (The negative 2 log-likelihood (− 2LL)) since the models were nested. Likelihood ratio test, Intra-class Correlation Coefficient (ICC), Median Odds Ratio (MOR), and Proportional Change in Variance (PCV) were computed to measure the variation between clusters. ICC quantifies the degree of heterogeneity of intimate partner violence between clusters (the proportion of the total observed individual variation in intimate partner violence that is attributable to between cluster variations).

ICC = ϭ²/ (ϭ² + π²/3) [35], but MOR is quantifying the variation or heterogeneity in outcomes between clusters and is defined as the median value of the odds ratio between the cluster at high risk of experiencing intimate partner violence and cluster at lower risk when randomly picking out two clusters (EAs) [36].

\(\partial\) ² indicates that cluster variance.

PCV measures the total variation attributed to individual-level factors and community-level factors in the multilevel model as compared to the null model.

Multilevel random intercept logistic regression was used to analyze factors associated with intimate partner violence at two levels to take into account the hierarchical structure of the data, at individual and community (cluster) levels. Four models were constructed for the multilevel logistic regression analysis. The first model (a multilevel random intercept logistic regression model without covariates) was an empty model without any explanatory variables, to determine the extent of cluster variation on intimate partner violence. The second model (determined the association between the individual level predictors and intimate partner violence) was adjusted with individual-level variables; the third model (determined the association between community-level variables and intimate partner violence) was adjusted for community-level variables while the fourth (individual and community level model) was fitted with both individual and community level variables simultaneously. The final model (a model with individual and community level factors) was chosen since it had the lowest deviance.

Variables with p-value ≤ 0.2 in the bi-variable analysis for both individual and community-level factors were fitted in the multivariable model. Adjusted Odds Ratio (AOR) with a 95% Confidence Interval (CI) and p-value < 0.05 in the multivariable model were used to declare significant predictors of intimate partner violence. Multi-collinearity was also checked using the variance inflation factor (VIF) which indicates that there is no multi-collinearity since all variables have VIF < 5 and tolerance greater than 0.1.

Permission for data access was obtained from major demographic and health survey through an online request from http://​www.​dhsprogram.​com. The data used for this study were publicly available with no personal identifier. We received the authorization letter from The Demographic and Health Surveys (DHS) Program. The IRB-approved procedures for DHS public-use datasets do not in any way allow respondents, households, or sample communities to be identified. There are no names of individuals or household addresses in the data files. The geographic identifiers only go down to the regional level (where regions are typically very large geographical areas encompassing several states/provinces). Each enumeration area (Primary Sampling Unit) has a PSU number in the data file, but the PSU numbers do not have any labels to indicate their names or locations. In surveys that collect GIS coordinates in the field, the coordinates are only for the enumeration area (EA) as a whole, and not for individual households, and the measured coordinates are randomly displaced within a large geographic area so that specific enumeration areas cannot be identified.

A total of 6090 women were included in the study. Of these, 3207 (52.7%) of women did not have formal education and 925 (15.6%) of the women attained secondary education or higher. About 1106 (18.2) of the women were from the poorest household and 2502 (41.1%) had media exposure. Nearly three-fourth (76.1%) of household heads were males and about 3580 (58.8%) of the women were participated in making decisions (Table 1). About 2281 (37.5%) and 1477 (24.3%) are living in the Oromia and Amhara regions, respectively. Regarding community media exposure and community poverty, about 47.3% of the women were from community with high media exposure and 51.9% from high community poverty (Table 2).

The overall national prevalence of IPV among reproductive-age women in Ethiopia was 33.5% (95% CI 32.1, 34.7). The spatial distribution of IPV was non-random in Ethiopia (Global Moran's I = 0.1, p-value < 0.0001) (Fig. 1). In the Getis Ord GI statistical analyses, the significant hotspot areas of IPV were located in the east SNNPRs, west Oromia, Gambella, north Amhara, and northwest Tigray regions, whereas significant cold spot areas of IPV were found in east Amhara, west Afar, and Somali regions (Fig. 2).

The SaTScan analysis identified a total of 192 significant clusters, of these 181 clusters were primary clusters located in Benishangul-Gumuz, Gambella, northwest Amhara and west Oromia regions centered at 10.637520 N, 35.719206 E with a radius of 373.97 km, a Relative Risk (RR) of 1.35 and a Log-Likelihood Ratio (LLR) of 16.55, at p < 0.001 (Table 3). This showed that women within the spatial window had 1.35 times higher risk of experiencing IPV than women outside the spatial window. The secondary clusters scanning window was located between the border area of the southwest Oromia, and north Tigray regions (Fig. 3). The Kriging interpolation identified northwest Tigray, northern and eastern Amhara, west Benishangul, east SNNPRs, and southwest Oromia regions as predicted high-risk areas of IPV while the Somali region was identified as predicted low prevalence of IPV (Fig. 4).