Cross-national and multilevel correlates of partner violence: an analysis of data from population-based surveys

Background On average, intimate partner violence aff ects nearly one in three women worldwide within their lifetime. But the distribution of partner violence is highly uneven, with a prevalence of less than 4% in the past 12 months in many high-income countries compared with at least 40% in some low-income settings. Little is known about the factors that drive the geographical distribution of partner violence or how macro-level factors might combine with individual-level factors to aff ect individual women’s risk of intimate partner violence. We aimed to assess the role that women’s status and other gender-related factors might have in defi ning levels of partner violence among settings.


Introduction
Violence against women by a male intimate partner is both a violation of women's human rights and a profound health problem that interferes with their full participation in society and their countries' social and economic development.
Although violence aff ects many women's lives, it does so unevenly. Research shows that the prevalence of violence diff ers greatly across settings-eg between countries, within countries, and across neighbourhoods and regions. The 12 month prevalence of partner violence (established with similar questions and methods between countries) varies from 4% in highincome countries such as Denmark, the UK, Ireland, and the USA to more than 40% of women in some lowincome countries such as Ethiopia. [1][2][3][4][5] In the WHO Multicountry Study on Women's Health and Domestic Violence (referred to as the WHO Study), reports of current abuse by a partner varied from less than 4% in Yokohama, Japan, and Belgrade, Serbia to 53·7% in e333 www.thelancet.com/lancetgh Vol 3 June 2015 rural Ethiopia and 34·2% the Peruvian department of Cuzco. 6 The average 12 month prevalence of partner violence across the 28 states of the European Union is likewise 4%. Even between neighbourhoods in a city or villages in a district, the prevalence of partner violence often varies substantially. 7 This fi nding raises a crucial question: what accounts for these diff erences in levels of violence and can the geographical distribution of violence yield insights useful for violence prevention?
Feminists have long contended that the main drivers of partner violence are gender-related norms and hierarchies that shape relationships between men and women and structure women's access to resources. 8 These factors, combined with genetic predispositions, developmental pathways, and partner-related and relationship-related factors, determine the likelihood that a couple will experience violence and drive the overall level of partner violence in a setting. Feministinformed theory acknowledges the role of individual lifecourse factors, but emphasises the importance of community and macro-level factors as fundamental in defi ning levels of abuse. 9 Research into intimate partner violence, however, has largely ignored the role of macro-level factors in aff ecting a woman's risk of violence and the geographical distribution of abuse. Violence research is dominated by studies from North America and other high-income settings and these have emphasised the role of personality and relationship dysfunction, childhood trauma and developmental adversity, and antisocial behaviour as key risk factors for partner violence. [10][11][12] Eff orts from US researchers to test the feminist hypothesis on the importance of gender norms and hierarchies at a state level have yielded equivocal results, 13 leading many academics to argue that gender plays a minor part in the cause of abuse. 14,15 Hence this study has two goals: to test the gender hypothesis by assessment of the degree to which macro-level factors related to women's status, gender inequalities, and norms of male authority and control are associated with population-levels of partner violence and to explore whether these factors interact with individual-level variables to predict a woman's personal risk of partner violence. Specifi cally, we examine the following four questions: do macro-level gender variables correlate with the geographical distribution of partner violence in the directions feminist-informed theory would suggest? What best accounts for the apparent association between a country's level of socioeconomic development and its overall prevalence of partner violence? Which factors remain important at the macro level when analysed in the presence of other macro-level and individual-level predictors of violence? Do important cross-level interactions exist between macro-level and individuallevel factors that aff ect a woman's personal risk of partner violence?
This analysis builds on and extends the fairly undeveloped scientifi c literature about macro-level predictors of population prevalence of violence against women. So far, only nine studies 13,[16][17][18][19][20][21][22][23] have sought to explore country-level or state-level predictors of partner violence and all have weaknesses in the methods they have used, especially with respect to the outcome variable utilised. One study 21 derives a numerical measure of partner violence on the basis of qualitative descriptions in human rights reports and the remainder rely on data from a range of studies that used diff erent defi nitions and measures of intimate partner violence. Our analysis is the fi rst to analyse macro-level predictors of partner violence at the level of the country and survey year with highly similar outcome data.

Research in context
Evidence before this study Before initiation of this study, we did a comprehensive, but non-systematic, review of the scientifi c literature on macro-level factors associated with partner violence. Between July 1, 2014, and August 8, 2014, we searched Econlit, JSTOR, Scopus, NBER Working Papers, Medline, and Global Health using the search terms: "macro*", "community*", "ecological", "determinant", "cross-national", "country-level", "neighbourhood", and various terms for partner violence (eg, domestic violence, wife abuse) and grey literature available on relevant websites. We searched only English language journals. Only 9 relevant studies were identifi ed, all with substantial fl aws in their methods.

Added value of this study
The current study is the fi rst to analyse macro-level predictors of partner violence with a well defi ned and highly similar measure of partner violence across countries, on the basis of self-reported victimisation in population-based surveys, all with the same questions, survey methods, and ethical controls. It shows that gender-related factors at the country level and regional level-especially norms and property rightspredict the population prevalence of current partner violence (physical or sexual violence in the past 12 months). The study also shows that the macro-environment can potentiate or dampen the eff ect that individual-level factors have on the risk of partner violence.

Implications of all the available evidence
Our fi ndings suggest that policy makers could reduce violence by elimination of gender bias in ownership rights and addressing norms that justify wife beating and male control of female behaviour. Prevention planners should place greater emphasis on policy reforms at the macro-level and take cross-level eff ects into account when designing interventions.  26 Prevalence surveys were selected for their similarity in terms of violence questions, methods, and ethical controls, on the basis of our knowledge of the area. Additionally, we used national-level statistics compiled by the UN, the World Bank, the Organisation for Economic Co-operation and Development (OECD), and topic-specifi c datasets compiled by academic institutions to track specifi c issues, such as women's economic and political rights. These institutions routinely obtain or make available country-level data for the economy, employment, education, health, and other national-level statistics compiled by governments.
Both the DHS and WHO studies use in-person household surveys to interview a representative sample of women aged from 15 years to 49 years, either nationally (in the case of the DHS and WHO surveys done in Samoa and Turkey) or subnationally in the remaining WHO surveys. Both surveys used behaviour-specifi c questions about diff erent acts of physical and sexual partner violence. Although wording about acts of violence diff ers slightly in some surveys, the variations are minor. All surveys used similar ethical guidelines designed to maximise safety and disclosure, including interviewing only one woman per household, maintaining complete privacy during the interview, and implementation of specialised sensitivity training for interviewers. 27,28

Outcomes
The outcome variable for this analysis is the population prevalence of current partner violence, defi ned as the percentage of ever-partnered women (excluding widows without a current partner), aged from 15 years to 49 years who experienced at least one act of physical or sexual violence within the past 12 months.
Our analysis focuses on partner violence in the past year to address diff erences in inclusion criteria between the DHS and WHO studies. The DHS is restricted to violence perpetrated by a woman's current or most recent partner, whereas the WHO study asks about violence perpetrated by any partner since the age of 15 years. By focusing on the previous 12 months for both surveys, we maximise similarity between them. Moreover, a comparison of how current macro-level factors aff ect present day rates of partner violence makes conceptual sense.
Our exposure variables represent various genderrelated domains and control variables that off er alternative explanations for the geographical distribution of violence. The gender-related domains include women's status, women's economic participation and entitlements, women's political participation and entitlements, gender inequality between men and women, and gender-related norms and attitudes. Additionally, we include variables to control for a country's level of socio-economic development (natural log of gross domestic product [GDP] in purchasing power parity in 2011 constant US dollars) and the age structure of the population. Table 1 summarises the individual data sources and variables used to represent each domain. All macro-level variables represent the mean level of that measure aggregated at the survey level (if derived from surveys) or a national-level measure, if taken from data banks maintained by multilateral agencies, such as the World Bank. Several of the indicators represent specialised indices of entitlements or discrimination created by academics or worldwide institutions to track gender-related trends. These include measures of women's political and economic rights from the Cingranelli-Richards Human Rights Database (eg, women's de jure and de facto economic entitlements) and two measures of gender inequality in family law and ownership rights created and maintained by the OECD as part of its Social Institutions and Gender Index (SIGI) database. In both indices, two experts independently assigned scores to countries on the basis of data from the US State Department's Country Reports on Human Rights Practices, according to a detailed coding scheme. The SIGI family law index, for example, assesses the degree to which states discriminate against women on issues of child guardianship and custody, access to divorce, the minimum legal age of marriage, and the right to inherit property. Values range from 0 (no discrimination between men and women in law and practice) to 1 (high discrimination between men and women).
For each explanatory variable tested, we used data from the same year that the violence survey was undertaken. Where an exact match was not available, the closest year to the survey date was used, giving priority to data obtained before the date of the violence survey.

Statistical analysis
This study uses various diff erent techniques to address our diff erent research questions. Scatterplots, histograms, and linear and quantile regression were used to assess normality, identify outliers, and examine the potential associations between macro-level explanatory variables and partner violence at a country and survey level. The goal of this bivariate ecological analysis was to assess whether the population-level distribution of partner violence is associated in the predicted direction with macro-level variations in women's status, gender inequality, and norms related to male authority and control. Quantile regression was used to check the robustness of our fi ndings. Because

For the US State Department's Country Reports on Human
Rights Practices see http:// www.state.gov/j/drl/rls/hrrpt/ For the SIGI database see http://genderindex.org/ quantile regression models the median rather than the mean, quantile regression generally yields more accurate coeffi cients for skewed datasets, with fewer covariates emerging as signifi cant. It also can be used to assess whether a covariate exerts a diff erential eff ect at low versus high ends of an outcome distribution. 29 Next, we ran the same ecological analysis with several variables. Standard errors are clustered at the country level to take into account that some countries have several surveys and therefore their observations are not fully independent. For this analysis, we used linear rather than quantile regression. Our strategy for model building was to establish which variable from each domain dominated when taken together with the other variables selected to represent that domain. We selected the most robust measure for each domain (highest, most stable eff ect size), and then ran a set of structured regressions to establish whether the apparent association between a country's aggregate GDP per person and its partner violence persisted in the presence of gender-related variables. All models include year-fi xed eff ects. Robust p values are provided in parentheses. We regarded p values less than 0·10 to be statistically signifi cant.
Multilevel analysis was used to examine whether the macro-level variables associated with the geographical distribution of partner violence were mainly a function of the characteristics of the individuals living there (a compositional eff ect) or suggestive of a higher order social process (a contextual eff ect). When we include the same variable at both levels, we essentially test whether there is an extra correlation between the macro factor and abuse in addition to that operating at the individual level.

Role of the funding source
The funder of the study had no role in the study design, data collection, data analyses, or data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had fi nal responsibility for the decision to submit for publication.   Table 2 summarises the bivariate associations between gender-related macro factors and the geographical distribution of partner violence. As indicated by the arrows, all the ecological associations are in the direction that feminist-informed theory would predict (with the exception of political rights) although some did not achieve statistical signifi cance, including early marriage, and the ratio of earned income between men and women. In addition to being associated with women's status and other gender-related variables, levels of partner violence seem lower in high-income countries than in low-income countries. For every log increase in GDP per person, the prevalence of partner violence decreases by 5·5%. These fi ndings are robust to quantile and logistic regression and are largely similar in urban and rural samples (appendix). Table 3 explores alternative explanations for why reductions in violence might accompany socioeconomic development. Each column represents a separate regression on intimate partner violence in the past 12 months. With model 1, the negative correlation coeffi cient for GDP per person (-0·055, p=0·009) confi rms that current partner violence decreases as the GDP increases. We postulated that GDP is actually a marker for more complex social processes and transformations in women's roles that frequently accompany economic growth and modernisation. Consistent with this theory, the correlation between GDP and partner violence decreases and becomes nonsignifi cant as we add in norms related to wife beating and male authority or control over women (models 2 and 3). We used models 4-6 to examine whether age structure, number of years in education, or the proportion of women working for cash could instead be responsible for the apparent association between norms and violence, but the ecological association remains statistically signifi cant in the presence of these additional controls

Table 3: Ecological analysis of macro-level factors related to gender and women's status on mean levels of current partner violence
See Online for appendix (table 3). Overall, this analysis suggests that the population prevalence of current partner violence against women is 14·6 percentage points higher in a setting where 100% of people agree with at least one of six justifi cations for wife beating (norms=1) compared with a setting where no one justifi es abuse (norms=0), all other things in the model being equal (model 7). With respect to the association between discrimination in women's ownership rights and current partner violence, the presence of ownership laws and practices that privilege men over women were robustly and signifi cantly associated with higher levels of violence (0·313, p=0·006). The SIGI ownership index seemed to be the strongest predictor of aggregate levels of partner violence of all variables in the gender inequality domain. When the index was broken down into its component parts, only women's access to land and other property were signifi cant, suggesting that ownership of assets rather than access to credit or banking drives the association (model 8). Additional analysis of urban versus rural samples further suggested that it is gender discrimination in ownership among women from rural areas that mostly accounts for the association present at a country level (appendix). Table 4 shows the results of the same analysis but with multilevel regressions. Multilevel coeffi cients represent the risk of partner violence to individual women in the presence of macro-level factors. The advantage of multilevel regressions compared with ecological associations is that they show how factors are important at diff erent levels in the social ecology. Table 4 shows that completion of secondary education (-0·080, p=0·007)     and being older than 34 years (-0·049, p=0·001) signifi cantly reduce a woman's personal risk of partner violence. Living in countries or regions where acceptance of wife beating and male authority are high remains signifi cantly associated with partner violence at the 10% level in the presence of these compositional controls. Table 4 shows that women who accept wife beating as a man's right and who have a controlling partner are at a signifi cantly higher risk of violence at the bivariate level (data not shown) and in multilevel analysis (0·046, p<0·0001 and 0·079, p<0·0001). Macro-level norms of acceptance and male authority are no longer signifi cant in the presence of these individual-level factors, but they might be on the causal pathway between norms and intimate partner violence. Table 4 also shows that living in a country that discriminates against women in access to land and other property remains a strong driver of abuse-related risk (0·132, p=0·015 and 0·155, p=0·003), even in the presence of a range of individual factors. Table 5 splits the sample between surveys in countries with a high acceptance of partner violence and a low acceptance of violence, showing potential cross-level interactions. High acceptance is defi ned as above the median of the survey mean of those accepting at least one justifi cation for wife beating. Table 5 shows that the level of overall acceptance of violence aff ects the eff ect of individual age-related and education-related variables on women's risk of partner violence. The coeffi cient for education is greater in settings with high acceptance compared with lower acceptance of violence, suggesting that education is more protective in countries or regions where justifi cation of wife beating is greater. Being in the age range of 15-24 years is also more risky in countries with high acceptance of violence (the p value of the diff erence is 0·064 for being 25-34 years compared with being younger than 25 years). All interaction p values are available in the appendix. Table 5 also splits the samples into surveys with very high and very low mean acceptance of wife beating, defi ned as being above the 80th percentile (where more than 48% of survey respondents accept violence) and below the 20th percentile (where less than 6% do). In countries with very low acceptance, a woman's education, age, and whether she works for cash make no diff erence to her risk of partner violence, but education and older age are protective in high acceptance settings. Individual acceptance of violence is much more strongly associated with being abused in areas where partner violence is highly normative than where it is not (p value of the diff erence is 0·004). This fi nding suggests an interaction between norms condoning violence and individual attitudes. By contrast, having a controlling partner seems slightly more dangerous in settings with very little acceptance of violence than in settings where partner violence is normative (0·090, p=0·007 compared with 0·070, p<0·0001).
We did the same analysis for surveys where many versus few women work (appendix). Working for cash increased a woman's risk of partner violence substantially more in settings where few women work than in settings where many women work (0·028, p<0·0001 in surveys in the lowest 20th percentile of women working vs 0·016, p=0·076 in surveys in the top 80th percentile of women working). Similarly, schooling is much more protective in settings with the lowest share of women working (bottom quintile) compared with the highest quintile (-0·130, p<0·0001 vs -0·042, p=0·073).

Discussion
Our analysis suggests that gender-related factors at the country and regional level help to predict the population prevalence of current partner violence (physical or sexual violence in the past 12 months). This includes factors related to women's status, such as educational achievement, women's access to cash or employment, and their de jure and de facto economic rights. Especially predictive of the geographical distribution of partner violence are norms related to male authority over female behaviour, norms justifying wife beating, and the extent to which law and practice disadvantage women compared with men in access to land, property, and other productive resources. Gender-related discrimination in family law, including diff erential rights to child custody, to inherit land and money, and to marry and divorce, also predict levels of partner violence across settings. Collectively, these associations provide suggestive empirical support for the gender hypothesis.
We similarly fi nd that despite the strong and consistent negative association between GDP per person and level of partner violence, level of socioeconomic development is unlikely to be causally related to prevalence of intimate partner violence. Rather, GDP per person seems to be a marker for other social processes that often accompany socioeconomic development. These include erosion of the belief in male superiority, entry of women into the paid labour force, and increased access to education and economic assets for women. More gender-equitable norms could naturally emerge as values shift from survival issues to greater emphasis on self-actualisation, individualism, and innovation, as modernisation theorists contend. 30 Alternatively, norms could shift in the face of women's emancipatory demands and widespread entry into the paid labour force. 31 Contrary to our expectations, partner violence was not associated with average prevalence of child marriage or gender inequities in the levels of secondary school completion or earned income. The tradition of child marriage might be restricted to specifi c regions or groups within a country and hence any association would be better captured at a community rather than at a national level. Previous research 32 has shown an association between child marriage and intimate partner violence at the individual level, but to our knowledge, no other studies have examined this association at an ecological level. With respect to secondary school completion, contrary to our expectations, rates were largely similar for boys and girls in many countries, making it a poor indicator of gender inequality. 33 Similarly, we suspect that reported levels of earned income are less reliable than data about employment or other economic indicators that were associated with intimate partner violence. These factors could partly account for the absence of an association. Alternatively, intimate partner violence could be more strongly associated with women's absolute status, rather than their relative status to men.
Our multilevel modelling suggests that macro-level processes aff ect women's individual risk of violence in addition to predicting the geographical distribution of abuse. Both gender norms and gender discrimination in access to land and property remain signifi cant at the macro level when adjusted for the age and educational level of the women living there. Macro-level norms become non-signifi cant when acceptance of violence and of a partner's controlling behaviour are added to the model; however, how to interpret this is unclear. Norms are likely to work precisely by aff ecting attitudes and behaviour, suggesting that these measures should not be in the regression because they are part of the causal pathway. As observed by Boyle and colleagues, 34 indiscriminately controlling for individual variables could attribute valid area-level eff ects to confounding when, in fact, they have set in motion person-level processes that increase risk of intimate partner violence. Additional research, including exploring norms at both the survey and cluster level, could help to clarify the situation.
Our stratifi ed analysis shows the importance of taking into account cross-level eff ects. A girl's education is more strongly associated with a reduced risk of partner violence in countries where wife abuse is normative than where it is not (as shown by the larger coeffi cient in the split samples with high acceptance). A similar statistical interaction exists between education and working for cash (at the individual level) and the overall proportion of women who work. Should the association prove causal, educating a girl would yield a bigger dividend in terms of reducing her risk of violence in countries where wife abuse is highly normative. At the ecological level, having many women in the formal work force is negatively associated with a country's level of partner violence, but at an individual level, where few women work, working for cash increases a woman's risk of partner violence. This helps to explain past confl icting fi ndings about the eff ect of employment on women's risk of violence. 35 These fi ndings hold insights for future programming to prevent partner violence in low-to-middle-income countries. First, greater emphasis must be placed on shifting normative expectations around the acceptability of wife beating and the perceived right of men to control female behaviour. Similarly, practitioners and researchers should explore removing barriers to women's access to land and property as a potential strategy for reduction of intimate partner violence levels. A study 36 of women in Kerala, India, identifi ed that women who own immovable property-especially a home-are at a substantially lower risk of both current and lifetime partner violence than are those who do not.
More generally, prevention planning must acknowledge that factors have diff erential eff ects at the macro, community, and individual level and that strong cross-level eff ects exist. Thus, a microfi nance or job-creation programme could increase a woman's risk of intimate partner violence in the short term, even though having many women in the workforce reduces a country's overall level of intimate partner violence. Similarly, some factors hold diff erential potential to reduce risk in high versus low violence settings, as shown by the larger coeffi cients in the quantile regressions run in countries with the highest levels of current partner violence compared with the lowest (top 20% of intimate partner violence distribution vs the bottom 20%). Increased understanding of these diff erentials could help better target prevention interventions. Given the potential of economic empowerment to increase violence in the short term, programmes must anticipate these risks and incorporate training for staff and safety planning with women to minimise any negative results of shifts in household gender dynamics.
Our fi ndings are only as valid as the reliability of the original data sources, some of which depend on government reporting. Because data for many of the World Bank and OECD exposure variables are available only for certain years, covariate and outcome data are not optimally time-matched for all countries. Because national-level indicators change slowly and explanatory variables aggregated from the studies are not subject to this concern, we do not regard this as a major threat to validity.
An inherent problem in all macro-level analyses is to separate correlation from causality. We do not claim causality for any of the correlations presented here. Many potential variables might aff ect both abuse and our exposures of interest. We do fi nd, however, that GDP is unlikely to be causally related to intimate partner violence, whereas norms and ownership rights are more likely to be. We urge future studies to use case studies and exploit natural experiments to disentangle the causal association between variables where possible.
Finally, we have used country or survey as our level of interest. Although this makes sense for factors such as laws and GDP, it might be too high a level to analyse acceptance and employment. Multilevel studies that have used clusterlevel or village-level data have identifi ed diff erent cross-level eff ects from the ones we identify in this study. For example, Boyle and colleagues 34 identifi ed that, in India, acceptance of violence at the neighbourhood or cluster level dampens the protective eff ect of education on violence, whereas we report that education is more protective in countries with a high acceptance of violence. Similarly, Cools and Kotsadam 37 fi nd that in Africa, women who work are at greater risk of abuse in clusters with higher acceptance of violence, but we identifi ed no such interaction at the macro level across our more geographically varied dataset. These diff erences show that the reduction of intimate partner violence needs attention to the variable eff ects the same factor might have at diff erent levels of the social ecology and a strategic matching of interventions to targeted level. We plan to explore community-level eff ects on intimate partner violence and cross-level eff ects of macro-level versus cluster-level factors in future analysis.

Contributors
LLH designed the study and developed the fi rst draft of the manuscript. AK did the analysis and contributed to the manuscript. Both authors coded the data and interpreted the results.

Declaration of interests
We declare no competing interests.