Data source
This is a cross-sectional analysis of population-based data from the 2018 NDHS. The NDHS is a nationally representative survey used to gather sociodemographic and other health-related indicators such as IPV [23]. A two-stage sampling procedure was used to gather data from 36 administrative units and the Federal Capital Territory (FCT). The survey's primary sampling unit was made up of samples drawn at random from clusters. A total of 41,821 women aged 15–49 participated in the 2018 study. From this number, 8968 women who had complete information and participated in the domestic violence module were considered in this study. The sampling, pretesting and the general methodology of the 2018 NDHS have been published elsewhere [24, 25]. In writing this manuscript, we adopted the guidelines for improving the reporting of observational studies in Epidemiology [26]. The dataset utilised in this study is available in the public domain and can be downloaded from https://dhsprogram.com/data/available-datasets.cfm.
Dependent variable
IPV was the outcome variable. It was obtained from the following variables: sexual violence, emotional violence, and physical violence. These three variables were derived from a series of questions in the domestic violence module that were related to a variety of violent acts that a woman had experienced. Previous studies include details on the questions for each aspect of three forms of IPV [27, 28]. There were Yes, or No response questions asked for each element of IPV. Therefore, a woman who had undergone at least one of the acts was regarded as ever experienced physical, emotional, or sexual abuse. From the questions asked on the experience of physical, emotional, and sexual abuse, IPV was created with respondents experiencing at least one of these violent acts regarded as ever had IPV and otherwise [9, 29].
Independent variables
Based on theoretical and practical significance and the availability of the variables in the dataset, we considered both individual and contextual factors in our study [23]. These were also influenced by their association with IPV in several previous studies in Nigeria and sub-Saharan (SSA) in general [9, 17].
-
a.
The individual-level factors were age (15–24, 25–34, 35+), educational level (No education, primary, secondary/higher), husband/partner’s educational level (No education, primary, secondary/higher), marital status (currently married, cohabiting, previously married), working status (not working, working), ethnicity (Hausa, Yoruba, Igbo, Others), religion (Christianity, Islam, Traditionalist & Others), parity (0, 1–3, 4+), and exposure to mass media (yes, no) [9, 17].
-
b.
The contextual factors were place of residence (urban and rural), wealth index (poorest, poorer, middle, richer, richest), region (North Central, North East, North West, South East, South South, South West), sex of household head (male, female), community literacy level (low, medium, high), community socioeconomic status (low, medium, high) [9, 17].
Analyses
We employed both spatial and multilevel analyses in analyzing the data.
Spatial analysis
Different statistical software like Excel, SaTScan, ArcGIS, and Stata 16 were utilized for spatial distribution of IPV in Nigeria. A total of 1400 clusters or Enumerations Areas (EAs) were considered for this study. Among these clusters, seven were dropped because they had no measured longitude and latitude data. The data were weighted with v005 (weighing variable) and geographic coordinate data were merged in Stata 16 and then exported to excel, which was finally imported to ArcGIS 10.7 for spatial analysis.
Spatial autocorrelation
To check whether there is clustering effect in IPV in Nigeria, spatial autocorrelation analysis was done. This analysis result gives Global Moran’s I value, Z-score and p-value for deciding whether the data is dispersed or random or clustered. Moran’s I value close to positive 1 indicates there is clustering effect, close to negative one indicates dispersed and close to zero indicates random. If p-value is significant and Moran’s I value is close to mean, that means IPV had clustering effect [30].
Hot spot
The hot spot analysis tool gives a Getis_Ord or Gi* statistics for cluster in the dataset. Statistical values like Z-score and p-value is computed to determine the statistical significance of the clusters. Results of the analysis with high GI* value means hot spot areas and low GI* value means cold spot areas [31].
Prediction of IPV
Spatial prediction is one of the techniques of furcating unsampled areas based on sampled areas. In Nigeria, a total of 1400 enumeration areas were selected to take a sample for this areas that believed to be representative of the country. A total of seven clusters had no enumeration longitude and latitude were dropped. Based on 1393 sampled areas, it is possible to predict the remaining parts of Nigeria. Ordinary Kriging prediction methods were used for this study to predict IPV in unobserved areas of Nigeria.
SaTScan analysis result
Bernoulli purely spatial model was applied to identify IPV clusters using 1393 enumeration areas. SatTscan Software was used for the analysis. First, the dataset was managed as appropriate for the SaTScan software. Women who faced IPV were taken as cases and women who did not face IPV were taken as controls. The Cluster number, longitude and latitude data were obtained from GPS dataset. The cluster size less than 50% of the population was taken as upper bound. A Monte Carlo replication was used for this study. Based on the above criteria, primary clusters were identified.
Statistical analysis
Multilevel analysis
A two-level multilevel binary logistic regression models were fitted to evaluate the individual and contextual (household and community level) factors linked to IPV experience among women in Nigeria. In the modelling, women were nested within households; then households were nested within clusters. To account for the unexplained variability at the community level, clusters were proposed as a random effect. A total of four models were fitted. Firstly, we fitted an empty model, model 0, which contained no predictors (random intercept). Following that, model I only included individual-level variables, model II only included contextual-level variables, and model III included both individual-level and contextual-level variables. The odds ratio and related 95% confidence intervals were provided for all models. These models were fitted using a Stata command “melogit” for the identification of predictors with the outcome variable (IPV). The log-likelihood ratio (LLR), Akaike Information Criteria (AIC) was used to compare models. The best fit model has the highest log-likelihood and the lowest AIC [32]. The multicollinearity test, which used the variance inflation factor (VIF), revealed no evidence of collinearity among the independent variables (Mean VIF = 1.83, Maximum VIF = 1.17, and Minimum VIF = 3.09). The domestic violence module sample weight (d005/1,000,000) was used in all analyses to account for over-and under-sampling, while the svy command was used to account for the complex survey design and generalizability of the results. All the analyses were carried out using Stata version 16.0 (Stata Corporation, College Station, TX, USA).
Ethical approval
Since the authors of this manuscript did not collect the data, we sought permission from the MEASURE DHS website and access to the data was provided after our intent for the request was assessed and approved on the 6th of April 2021. The DHS surveys are ethically accepted by the ORC Macro Inc. Ethics Committee and the Ethics Boards of partner organizations in different countries, such as the Ministries of Health. All methods were performed in accordance with the relevant guidelines and regulations. The women who were interviewed gave informed consent during each of the surveys.