Study design and data collection
This was a cross-sectional study that used data from the population-based 2013 Nigerian Demographic and Heath Survey (DHS). The DHS collected data from February – June 2013, via a stratified three-stage cluster sample design using a sampling frame containing the list of enumeration areas prepared for the 2006 Population Census of the Federal Republic of Nigeria [6]. Contiguous enumeration areas were joined to make a DHS cluster (primary sampling unit [PSU] representing one community each). The sampling yielded 904 PSU and 40,320 households from rural and urban areas [6].
However, we used only 896 PSUs in our analysis, as these were the ones covered for the IPV data. Each PSU had approximately 42 observations. A minimum of 30 observations per group, and 30 groups at the second level of the analysis is recommended [46, 47]. For cross level interactions, a minimum of 20 observations per group and a minimum of 50 groups is recommended [48], while 200 groups with minimum 20 observation per group is recommended if the slope variance is estimated [49]. Increasing the number of PSUs will yield more precise estimates of community effects than increasing the number of people within the PSUs [47].
Trained DHS field interviewers speaking the same language as respondents collected data using questionnaires by face-to-face interviews. Women aged 15–49 years in each household were eligible for interview. Also, a subsample of one eligible woman per household was randomly selected to be asked additional questions regarding domestic violence. Where there was more than one eligible woman in a household, the DHS used the Kish grid to select one woman [6]. Furthermore, in every second household, all men aged 15–49 years who were either permanent residents of the households or visitors present in the households on the night before the survey were eligible to be interviewed. Men were interviewed using a questionnaire that was similar to, but shorter than the women’s questionnaire. Details of the survey design and sampling procedure are discussed elsewhere [6].
Of the 39,948 women who participated in the survey, 27,749 (69.4%) were randomly selected to be interviewed for the domestic violence module. Given that the present study focused on IPV, 6745 (24.3%) women were excluded, as they had never been in a relationship. Thus, data for this study were based on 21,004 ever-partnered women. IPV was assessed in the DHS based on a modified, shortened, and previously validated version of the Conflict Tactics Scale (CTS) [50]. In total, 202 (0.96%) women were excluded due to missing data of one or more variables, bringing the final number to 20,802 women in 896 PSU. A total of 17,359 men were interviewed, however only data from 17,194 men were analysed in this study due to missing data of 165 (0.95%).
Study setting
The presence of 374 ethnic groups in Nigeria’s 36 states mean that cultural practices and gender norms differ [51, 52]. The Tiv-speaking people of North Central Nigeria believe wife-beating is a sign of affection and love [53]. Among the Igbos in South Eastern Nigeria societal privileges such as traditional titles, lands, wealth and decision-making are male-centered, and exclude women [51, 52]. Marriage customs in the largely patriarchal society of Nigeria involves payment of bride price. This practice often gives men an excuse to lay claims to ownership of their wives [7, 52].
In 1984, Nigeria became a signatory to the CEDAW, and ratified it in 1985 [6], but this has done little to protect women from discrimination and violence due to the long and laborious process of enforcing it [54,55,56]. The Nigerian criminal code makes provision for punishing unlawful and indecent assault on women, girls and men; three years imprisonment for assault on men, while assault on women and girls is punishable with two years imprisonment [57]. There seems to be a contradiction, however, as the Penal code, which governs the states in Northern Nigeria, allows husbands to “correct their wives using physical punishment, so long as the woman is not seriously harmed”. Furthermore, cases of domestic violence in Nigeria are hardly ever brought to trial as law enforcement agents consider domestic violence to be family affairs which should be resolved within the family. Particularly in rural areas, police do not respond if they consider the cases to be within cultural norms [58]. It is challenging to harmonize legislation and eradicate discriminatory measures due to the concurrent implementation of civil, customary and religious laws which sometimes contradict each other [59].
Operationalisation of variables
Outcome variable
IPV as the outcome of interest was measured as physical violence, sexual violence, and emotional violence. Questions included experiences of one or several of the following acts of abuse by a current or former partner:
-
Physical violence: i) pushing, shaking, or throwing something at her; ii) slapping her; iii) twisting her arm or pulling her hair; iv) punching her with his fist or hitting her with something harmful; v) kicking, dragging, or beating her; vi) choking or burning her on purpose; and vii) threatening or attacking her with a weapon (e.g., gun or knife).
-
Sexual violence: viii) forced sexual intercourse; ix) physically forcing her to perform any other sexual act when undesired; and x) forcing her with threats to perform sexual acts when undesired.
-
Emotional violence: xi) humiliating her in public; xii) threatening to hurt or harm someone close to her; and xiii) insulting or making her feel bad about herself.
Physical violence had a Cronbach’s alpha of 0.82; sexual violence by items viii–x, α = 0.84; emotional violence by items, α = 0.74; and any physical, sexual, or emotional violence, α = 0.87, indicating overall good test performance of the interview questions. A respondent was considered to have experienced IPV if she answered yes to at least one act of any of the forms of violence (physical, sexual, or emotional).
Exposure variables
Individual-level factors
Women's status –Women's status/empowerment encompasses several dimensions of a woman’s life – socio-cultural, economic, familial/interpersonal, political, legal -, at various levels – individual, family/household, community and the larger society. Practically, it is not so easy to separate these dimensions, as they may overlap considerably [18, 20]; thus, for the present study variables were selected that function as proxies for different dimensions [18, 20, 21]. Eleven single items were used to create a women's status index via principal component factor analysis. The Kaiser–Meyer–Olkin Measure of Sampling Adequacy (KMO) was 0.81, indicating that the variables used in creating the index were adequate for principal component analysis.
Employment status and earnings: Respondents were asked whether they were working, to which they responded ‘yes’ or ‘no’. Those who answered ‘yes’ were further asked what kind of earnings they got from their jobs; responses were ‘no earnings’, ‘cash earnings’, and ‘earnings in kind’.
Control of income: Respondents were asked who decided on how their earnings were used. Responses were categorised as ‘no earnings’, ‘has no control over her earnings’, and ‘decides solely or jointly with her partner’.
Education was categorised as illiterate, primary, secondary, or higher. Media exposure: An ordinal variable created from responses to three individual questions about how often a respondent read newspapers, listened to radio, or watched television. Responses were categorised as ‘no exposure’, ‘less than weekly exposure’, or ‘weekly exposure or more’.
Age at first marriage/cohabitation defined as median age in years when women aged 15–49 first married or lived with consensual partner [21]. This variable was categorised into four age groups: ‘less than 18’, ‘18–24’, ‘25–31’, and ‘more than 31’.
Participation in household decision-making measures women’s participation in the following items: who decides on the woman’s healthcare, who decides on large household purchases, and who decides on visits to relatives. For each item, a woman participates in decision-making when she alone or jointly with someone makes the decision. Responses are categorized as “no participation” or “participation”. These items reflect the degree of decision-making that a woman can exercise in areas that affect her own life and environment [21].
Eleven items were analysed for the index, from which 11 factors were generated; each factor, corresponding to one item. The Kaiser criterion, a rule of thumb, was used to determine the number of factors to be retained. Based on this criterion, factors with eigenvalues greater than 1 were retained, leading to three factors being retained [60]. These three factors explained about 72% of the total variability in the original 11 items. Some scholars have argued that the Kaiser criterion could lead to overestimation in the number of factors extracted [61], thus, the scree plot (see Additional file 1) was also used in conjunction with the Kaiser criterion to determine the number of factors to retain. The ideal pattern of the scree plot is a steep curve followed by a bend (elbow), which then begins to flatten out. The number of factors to be retained is the data points above the bend [60].
Furthermore, we examined the loadings of the three factors retained, on each of the original 11 items used in the analysis (see Additional file 2). The factor loading for a variable is a measure of how much the variable contributes to the factor; thus, high factor loading scores indicate that the dimensions of the factors are better accounted for by the variables [60]. A general rule is that for larger sample size, smaller loadings are allowed for a factor to be considered significant [62]. For a sample size of at least 300, a rotated factor loading of 0.32 is needed for the factor to be considered statistically meaningful [60, 61]. Items 1–3 load highest on factor 1. These items representing employment, income and control of income, correspond to the economic dimension of women’s status. Items 4–8 loads highest on factor 2, and represent exposure to newspaper, radio and television, education level and age at first cohabitation/marriage. These correspond to the social dimension of women’s status. Factor 3 has the highest correlation with items 9–11, which represent participation in household decision-making and correspond to the familial dimension. The uniqueness is the proportion of variation in an item not explained by a factor. Values more than 0.6 are usually considered high, which means that variable is not well explained by the factors [60].
The analysis predicted an index score for each woman, which was categorized into tertiles – low, middle and high status.
Covariates
Socio-demographic characteristics: i) age (four categories, 15 to 49 years), ii) place of residence (urban or rural), and iii) household wealth (categorised by the DHS into quintiles). Details of the wealth quintiles creation can be found elsewhere [63].
Attitude towards wife-beating – a categorical ‘yes’ or ‘no’ variable was created from responses to five scenarios: if she goes out without telling him; if she neglects the children; if she argues with him; if she refuses to have sex with him; and if she burns the food. An answer of ‘yes’ to at least one scenario meant the respondent justified wife-beating and was coded as 1, while an answer of no in all scenarios meant the respondent did not justify wife-beating and was coded as 0. Cronbach’s alpha of reliability calculated in this study for the items was 0.89. The above-mentioned five scenarios were chosen based on prevailing socio-cultural gender norms relations (6,40). Patriarchal societies are characterised by power relations and men’s authority over women. In these societies, women are expected to care for children, prepare food properly, keep the house clean, attend to husband’s sexual need, obtain husband’s permission before going out, be submissive to husband. Transgression of these expectations could be a trigger for wife-beating in a bid to discipline the woman [14, 15, 52].
Partner’s controlling behaviour – a binary ‘yes’ or ‘no’ variable – was derived from responses to five items: jealous if she talks with other men; accuses her of unfaithfulness; does not permit her to meet her friends; tries to limit her contact with family; and insists on knowing where she is always. Women who responded ‘yes’ to one or more questions were categorised as having a partner/husband with control issues. Women who responded ‘no’ to all the questions were categorised as not having a partner with control issues. This was based on only women’s responses. Cronbach’s alpha for this item was 0.90. The DHS included these series of questions to assess the degree of control exercised by a husband/partner over the respondent. An important early warning sign of violence in a relationship is control and close monitoring of women by their husbands/partners [6].
Contextual factors
i) Community norms about wife-beating was created by aggregating responses from men in each community. Men were asked if wife-beating was justified in the following scenarios: if she goes out without telling him; if she neglects the children; if she argues with him; if she refuses to have sex with him; and if she burns the food. Communities were categorised as ‘does not justify wife-beating’ if the proportion of men was 0% and ‘justifies wife-beating’ if the proportion of men that justified wife-beating was above 0%. ii) Control over female behaviour was created by aggregating women’s responses about their partner’s controlling behaviour in each community. Communities were grouped into tertiles of low, moderate, and high levels of control over female behaviour.
Statistical analyses
Descriptive analysis was conducted to present the proportion of women who experienced any IPV for each category in the explanatory variables. To compensate for non-response rates and women’s unequal selection probability, sampling weights (DHS domestic violence weights) were introduced in the descriptive statistics, and the results of the descriptive analysis were presented as numbers and weighted percentages. Bivariate analysis was performed via simple logistic regression to assess the association between individual women characteristics and IPV. The significance level was set at p-value = 0.05. Due to the hierarchical structure of the data, where individuals are nested within PSU (communities), a multiple multilevel logistic model [64, 65] with two levels (individual and community) was fitted to assess the effects of measured individual- and community-level (fixed effect) characteristics on women’s experience of IPV, and to estimate the extent of variations across communities (random effects).
Six models were fitted: A null model with no explanatory variables was used to show variation across communities and to justify the use of multilevel analyses. Model 2 contained only individual variables, showing random intercepts and fixed slopes. This model studied the association between women's status and IPV, adjusting for other potential confounders in the association and showed how much of the variation in IPV across communities was explained by individual-level factors. Model 3 was like model 2, but also contained community variables to show measures of association and to quantify how much community-level factors explained the IPV over and above individual-level factors. Model 4 was a random intercept random slope model, with individual variables only. We assumed that the effect of women status on IPV might be different from one community to another. In that case, the slope of the association between women’s status and IPV would vary from one community to another and community disparities become a function of individual women's status. Model 5 was like model 4 but included community-level variables. Each community had its own coefficient for the association between individual women’s status and IPV exposure. The random slope analyses provided information about whether the association between women status and IPV differed across communities to ultimately justify the examination of cross-level interaction. Model 6 was a full model that included a term for interaction between individual women’s status and men’s justification of IPV at the community level. We tested for only one cross level interaction.
Fixed effects (measures of association)
The results were expressed as odds ratios (OR) with 95% confidence intervals (CI). Statistical significance was determined at p-value < 0.05.
Random effects (measures of variation)
We calculated the second-level variance (variation between communities) regarding the prevalence of IPV (i.e., the intercepts in the multilevel logistic regression) and the second-level variance regarding the association between women status and experience of IPV (i.e., the slope variance in the multilevel regression). The slope variance tells us how each community’s coefficient for the association between women's status and IPV deviates from the population average. We also calculated the covariance between intercept and slope residuals. The covariance gives information about whether the association between individual women status and IPV depends on the community norms regarding IPV in the different communities (i.e., cross-level interaction). We also applied the intra-class correlation (ICC) and median odds ratio (MOR) to test the hypothesised phenomenon that individuals living in the same community shared a similar probability of experiencing IPV, after adjusting for the individual characteristics studied. The ICC gives us the proportion of the total variation at the community level, while the MOR expresses the community variance in the OR scale. If the MOR is equal to 1 (no community-level variation), there is no difference between the communities regarding IPV. The higher the MOR, the more important the contextual effects for understanding the individual probability of experiencing IPV.
The model fit was analysed using deviance information criterion (DIC) as a measure of how well our different models fitted the data. A lower value in DIC indicates a better fit of the model [66]. Parameters in the model were estimated using the mean–variance adaptive Gauss–Hermite. The Stata Version 14.1 (Stata Corp. Inc., TX, USA) software package was used for the analyses.
Ethical consideration
The survey procedure and instruments used in the DHS had already received ethical approval from the National Health Research Ethics Committee of the Federal Ministry of Health of Nigeria and the Ethics Committee of the Opinion Research Corporation Macro International, Inc. (ORC Macro Inc., Calverton, MD, USA). In line with WHO recommendation, only one woman per household was interviewed so that no one else in the household knew which issues were discussed. Interviewers reiterated informed consent immediately prior to administering domestic violence questions. Care was taken by interviewers to ensure privacy; where this was not possible, the interview was not conducted, or it was terminated if privacy was breached [6]. Permission to use the DHS data in the present study was obtained from ORC Macro Inc. The dataset does not contain any individual identifiers that would make it possible to track any participant.