Social Science & Medicine 69 (2009) 691697

Contents lists available at ScienceDirect

Social Science & Medicine

journal homepage: www.elsevier.com/locate/socscimed

Community inuences on intimate partner violence in India: Womens education, attitudes towards mistreatment and standards of livingq
Michael H. Boyle*, Katholiki Georgiades, John Cullen, Yvonne Racine
McMcaster University Hamilton, Ont. Canada

a r t i c l e i n f o
Article history: Available online 18 July 2009 Keywords: India Intimate partner violence Education Clustering Surveys

a b s t r a c t
Intimate partner violence (IPV) directed towards women is a serious public health problem. Womens education may offer protection against IPV, but uncertainty exists over how it might reduce risk for IPV at the community and individual levels. The objectives of this study are to: (1) disentangle community from individual-level inuences of womens education on risk for IPV; (2) quantify the moderating inuence of communities on individual-level associations between womens education and IPV; (3) determine if womens attitudes towards mistreatment and living standards at the community and individual levels account for the protective inuence of womens education; and (4) determine if the protective inuence of education against IPV is muted among women living in communities exhibiting attitudes more accepting of mistreatment. Study information came from 68,466 married female participants in the National Family Health Survey conducted throughout India in 19981999. Multilevel logistic regression was used to address the study objectives. IPV showed substantial clustering at both the state (10.2%) and community levels (11.5%). At the individual level, there was a strong non-linear association between womens education and IPV, partially accounted for by household living standards. The strength of association between womens education and IPV varied from one community to the next with evidence that the acceptance of mistreatment at the community level mutes the protective inuence of higher education. Furthermore, womens attitudes towards mistreatment and their standards of living accounted for community-level associations between womens education and IPV. Place of residence accounted for substantial variation in risk of IPV and also modied individual-level associations between IPV and womens education. At the community level, womens education appeared to exert much of its protective inuence by altering population attitudes towards the acceptability of mistreatment. However, there was no residual association between womens education and IPV at the community level once living standards are taken into account. While womens education provides strong, independent leverage for reducing the risk of IPV, planners must keep in mind important community factors that modify its protective inuence. 2009 Elsevier Ltd. All rights reserved.

Intimate partner violence (IPV) is the intentional use of physical force (beatings, rape) to inict harm on a spouse or partner. It is a widespread public health problem with serious consequences (Ellsberg, 2006). A recent investigation of Womens Health and

q Michael Boyle is supported by a Canada Research Chair in the Social Determinants of Child Health. Katholiki Georgiades is supported by an Ontario Mental Health Foundation New Investigator Fellowship. The authors thank Jon Rasbash for his helpful comments on the manuscript. * Corresponding author. Department of Psychiatry and Behavioural Neurosciences & Offord Centre for Child Studies, Hamilton Health Sciences, McMcaster University, Chedoke Site, Central Building 303, Hamilton, Ont., Canada L8N 3Z5. Tel.: 1 905 521 2100; fax: 1 905 521 4970. E-mail address: boylem@mcmaster.ca (M.H. Boyle).
0277-9536/$ see front matter 2009 Elsevier Ltd. All rights reserved. doi:10.1016/j.socscimed.2009.06.039

Domestic Violence against Women estimated the lifetime prevalence of IPV directed towards women to vary from 15 to 71% in 15 sites from 10 countries (Garcia-Moreno, Jansen, Ellsberg, Heise, & Watts, 2005). The physical and mental health sequelae of IPV directed towards women include increased mortality, injury and disability, worse general health, chronic pain, reproductive disorders, depression, PTSD, alcohol abuse and drug abuse (Golding, 1999; Plichta, 2004). Given that IPV is a product of social context (Jewkes, 2002), it is not at all clear that traditional medical approaches such as individual screening and intervention (MacMillan et al., 2006) will represent effective or efcient strategies for reducing IPV. An alternative is to focus on the social determinants of IPV with a view to identifying modiable characteristics for prevention.

692

M.H. Boyle et al. / Social Science & Medicine 69 (2009) 691697

Violence directed towards women in India has attracted special attention (Sharma, 2005). Historically, patriarchal prerogatives have dominated family relationships in that country and culturally sanctioned use of physical violence by men against women has been condoned as a mechanism of control and conict resolution (Segal, 1999; WHO, 2002). In addition, socio-economic inuences that elevate risk for IPV directed towards women poverty (Bangdiwala et al., 2004; Hindin & Adair, 2002; Jewkes, 2002; Koenig, Stephenson, Ahmed, Jejeebhoy, & Campbell, 2006), young age at marriage, low education (Bangdiwala et al., 2004; Gage, 2005; Jewkes, 2002; McQuestion, 2003) and multiparity (McQuestion, 2003) have been endemic to India in the past. India, however, contains substantial diversity in language, culture, socio-economic well being and demography (Dyson & Moore, 1983; Haub & Sharma, 2006). These factors can be expected to inuence normative expectations for gender roles, relations between men and women and the acceptability of using physical violence, providing a basis for area-patterning of IPV. Although mostly rural, India is also an emerging economic power and subject to forces of modernity, including increased levels of literacy and education for both men and women. These forces are most likely to be concentrated in selected geographic areas and also contribute to spatial differentiation in the risk for IPV. Womens education in India and risk for IPV The second National Family Health Survey (NFHS-2) conducted throughout India in 19981999 (International Institute for Population Sciences (IIPS) and ORC Macro, 2000) reported a strong negative gradient between educational attainment for married women and their reports of being beaten or physically mistreated in the past 12 months (recent IPV): illiterate, 14.1%; <middle school, 8.8%; middle school complete, 7.0%; and high school complete and above 3.6%. In a recent study of the same data set, these graded individual-level effects persisted after adjustment for a number of covariates (Ackerson, Kawachi, Barbeau, & Subramanian, 2008). Demonstrating contextual-level effects for womens education aggregated to the area level must explain between-area variation in IPV after controlling for potential confounding variables, including individual education. The same report by Ackerson et al. (2008) estimated that 6% of the total variation in recent IPV was attributable to male and female literacy aggregated to the neighbourhood level. In contrast, two studies, one in Bangladesh (Koenig, Ahmed, Hossain, & Mozumber, 2003) and the other in north India (Koenig et al., 2006) found no empirical evidence to support contextuallevel inuence of womens education on IPV. These negative ndings could be due to the lack of IPV clustering no statistically signicant place-to-place variations in IPV were reported and/or the dominance of other individual and community-level variables included in the regression models. Mechanisms of effect at individual and community levels

personal capability and opportunities for employment, as well as improved socio-economic circumstances and lower risk of IPV through the process of assortive mating. At a theoretical level, there are reasons to believe that womens educational attainment aggregated to the community level might exert protective inuences on womens risk for IPV. Higher education is aligned with more liberal norms and values pertaining to womens rights and less acceptance of violence as a means of resolving conicts. As more and more women in a particular area are exposed to higher education, the percent of the population embracing liberal views about gender equality and opportunity should increase. The increasing presence of these norms, values and attitudes may take on a collective inuence, lowering community tolerance for physical mistreatment and perhaps leading to sanctions against such behavior. This collective inuence would be based on processes associated with social learning and imitation and emerge from repeated social interactions and exchanges that come to dene the parameters of acceptable behavior (Kravdal, 2004). In addition to the increased presence of liberal views about gender equality and opportunity, we can expect increases in womens education at the community level to be associated with improved living standards, population health, public infrastructure and institutional capacity for serving the public good. It is possible that the effects of womens education on risk for IPV are indistinguishable from those associated with community socio-economic circumstances. Indeed, areas with higher living standards may provide greater opportunities for womens education. Demonstrating an independent effect for womens education aggregated to the community level would provide compelling evidence for its protective inuence. Community-level modication of individual-level effects Just as communities may exert a direct inuence on risk for IPV, they may also mute or augment risk or protection operating at the individual level. A person-environment t framework lies behind these community-level inuences that become manifest in crosslevel interactions. Serving to illustrate this is the reported interaction between womens education and neighbourhood literacy in the study by Ackerson et al. (2008). The protective effects of education were strongest for educated women (>6 years) living in high literacy neighbourhoods but weaker for women with levels of education discordant with their neighbourhood. In the present study, we associate higher levels of educational attainment with increased knowledge, an enhanced capacity to access and to use information, more autonomy and more liberal ideas about the status of women (Jewkes, 2002). The inuence of these individuallevel characteristics on IPV could be muted or even reversed in circumstances where regressive ideas about the status and role of women are dominant (Sugarman & Frankel, 1996). Objectives

Womens education is at the intersection of many inuences. Educational opportunities for girls will be subject to: 1) parental resources, beliefs and attitudes; 2) accessibility of schools and educational resources; and 3) community norms and expectations. Education can be expected to have an indirect effect on the risk of IPV by inuencing the beliefs, self-image and capability of young women. This should lead to changes in womens attitudes towards, and acceptance of, physical mistreatment as a means of subjugation and conict resolution: reducing individual tolerance for such behavior should lower risk of exposure. Of course, the by-products of education for women go beyond changes in attitudes; they include higher living standards that come from increases in

This study examines area-patterning of IPV in India with a special focus on womens education, attitudes towards mistreatment and living standards assessed at the community level and individual levels. The objectives of the study are to: (1) disentangle community from individual-level inuences of womens education on risk for IPV; (2) quantify the moderating inuence of communities on individual-level associations between womens education and IPV; (3) determine if womens attitudes towards mistreatment and living standards at the community and individual levels account for community-level associations between IPV and womens education; and (4) determine if the protective inuence of

M.H. Boyle et al. / Social Science & Medicine 69 (2009) 691697

693

education against IPV is muted among women living in areas exhibiting attitudes more accepting of mistreatment. The study builds on the work of Ackerson et al. (2008) by extending our understanding of the factors at the community and individual levels associated with education which might explain its protective inuence on IPV. Methods In the NFHS-2, female interviewers used standard survey questionnaires administered face-to-face to collect health-related information from a nationally representative probability sample of women aged 1549 years living in household dwellings. The survey used a stratied, multi-stage, cluster design based on the 1991 Census. Each state was divided into urban and rural areas. In urban areas, wards listed in the 1991 Census were stratied by district and female literacy, and a sample of wards was selected with probability proportional to size (PPS) (stage 1). Next, selected wards were divided into census enumeration blocks (CEBs) of about 150200 households each and one CEB was taken from each ward with PPS (stage 2). After listing households within selected CEBs, about 30 households were chosen with equal probability for enlistment in the study (stage 3). In rural areas, villages listed in the 1991 Census were stratied by district and several demographic features. Contiguous villages with less than 50 households were linked to form primary sampling units (PSUs) with more than 50 households, while larger villages, with more than 500 households, were segmented into smaller PSUs. Next, villages and PSUs were selected with PPS (stage 1). After listing households within selected villages and PSUs, about 30 households were selected with equal probability for enlistment in the study (stage 2). To our knowledge, none of the spatial areas selected for the NFHS-2 were contiguous giving us little reason to be concerned about spatial autocorrelation. Sample for analysis A total of 91,196 households were enlisted (97.5% response) and 89,199 women completed an interview (95.5% response). To be eligible for the analysis, women had to be usual residents in the household as well as married and living with their spouse at the time of interview (N 79,160). In households with more than one woman, one was selected at random to prevent household clustering from inating area estimates (N 69,750). (There were too few women in households to model household clustering as a separate level.) Also, women had to have complete information on the study variables to be included (N 69,191) and areas with fewer than 10 eligible women in the study were dropped from the analysis to reduce statistical overlap between variables measured on individual women and the same variables aggregated to the area level (N 68,466). Concepts and measures Intimate Partner Violence (IPV) Three questions taken from the Status of Women module were used to classify IPV in the last 12 months. The stem question read, Since you completed 15 years of age, have you been beaten or mistreated physically by any person? Women responding yes were asked to identify their relationship to all persons responsible for such acts and then to report the aggregate frequency of occurrence in the last 12 months in three categories: once, a few times and many times. Women answering yes to the stem question,

identifying their husband as the perpetrator and reporting an occurrence in the past 12 months were classied as positive for IPV. Independent variables These included womens educational level, household standard of living and womens attitudes towards acceptance of partner mistreatment. Womans educational level was measured by the total number of years in school as reported by the woman. Household standard of living was measured by a cumulative weighted index of 30 durable goods developed to assess the standard of living for households participating in the NFHS-2. The scale scores were converted to a standard score with mean 0.0 and standard deviation of 1.0. Coefcient alpha for the scale is 0.85. The approach is similar to that used by Filmer and Pritchett (1999, 2001) to create the asset index which performs as well as consumption expenditures (Bollen, Glanville, & Stecklov, 2001; Houweling, Kunst, & Mackenbach, 2003). Acceptance of partner mistreatment was measured by a six-item index comprised of binary indicators coded 0 (absent) versus 1 (present) for a woman who agreed that a husband is justied in beating his wife under one or more of the following circumstances: 1, if he suspects her of being unfaithful; 2, if her natal family does not give expected money, jewellery, or other items; 3, if she shows disrespect for in-laws; 4, if she goes out without telling him; 5, if she neglects the house or children; and 6, if she doesnt cook food properly. This index is a linear combination (the largest principal component) of the 6 indicator variables weighted by componentscore weights that maximally account for their co-variation. The rst principal component accounted for 52% of the variance and coefcient alpha for the scale was 0.82. Control variables The individual control variables included: womans age in years, family structure, total number of children, working outside the home, exposure to physical mistreatment (other than IPV) since age 15 years and urban residency. The stem question used to code IPV provided the basis to classify exposure to violence other than IPV. Women were classied positive for exposure since age 15 if they identied one or more perpetrators from the following groups: immediate family (mother, father, step mother, step father, son, daughter, brother or sister); extended family (mother-in-law, father-in-law, son-in-law, daughter-in-law, brother-in-law, sisterin-law, other relative); and other (friend, acquaintance, teacher, employer, stranger). Family structure was coded 0 (extended) versus 1 (nuclear). Extended families included the presence of parents or in-laws, whereas nuclear families consisted of a couple with or without children in the home. Urban residency coded 1 (yes) versus 0 (no) was assessed at the community level. Community measures Because India has relatively few states (n 26) in contrast to large numbers of small areas or communities (PSUs in this study), we focus on the latter for estimating specic area-level inuences. The community-area measures included: womens education, household standard of living and attitudes acceptant of partner mistreatment. These community-level measures were derived by summing values obtained on individual women in each community and dividing by the total number of women respondents living in each one. The aggregation of individual-level measures to represent group characteristics is recommended for testing ecological and individual-level hypotheses in multilevel studies (Blakely & Woodward, 2000). On average, these ecological measures are based

694

M.H. Boyle et al. / Social Science & Medicine 69 (2009) 691697

on 22 women per community (from 10 to 59) a sufcient number to generate reliable estimates. Analysis This study uses multilevel logistic regression analysis to model IPV reported by women. In logistic regression analysis, the probability of response (p) is converted to an odds: p/(1-p); and then transformed to a logit (p): ln [p/(1-p)] so that expected response values can be expressed as a linear function of the explanatory variables in the logit scale. Multilevel logistic regression analysis is used to model binary response variables that are nested hierarchically. In this study, reports of IPV are nested in i women, from j communities, located in k states. Multilevel analysis provides the basis to model correlated responses by partitioning residual errors associated with each level in the hierarchy and expressing these errors as patterns of variation (random effects variances). These random effects variances are used to estimate intra-class correlation coefcients (ICC) which quantify the extent of clustering at a particular level. The b coefcients in multilevel logistic regression analysis are called xed effects and quantify the strength of association between dependent and independent variables. Fixed effects regressions modelled at one level (e.g., individual) can be allowed to vary randomly at higher levels, in a way that is analogous to generating separate regressions for each group and then estimating their variability. For example, specifying the regression of IPV on womans education as a random effect at the state level would produce an estimate of how much the b coefcients for womens education vary from one state to another. The multi-level model used in the analyses is depicted below:

term between womens attitudes towards mistreatment assessed at the community level and womens education assessed individually. In preparation for the analysis, we centered the following variables (subtracted the overall means from their observed values): womens education at the community and individual levels; womens age in years and total number of children. Rescaling provides more realistic modelled estimates for the average levels of IPV and reduces problems of multi-collinearity when testing for statistical interactions. Results Table 1 presents summary information on the total sample. There are 26 states, 3118 areas (clusters) and 68,466 women. There is substantial variation in the prevalence of IPV between states (1.5 to 19.0%) and between clusters (0.0 to 61.5%). About 31.3% of women live in urban areas and 51.4% live in nuclear families. On average, women have 3.90 years of education and 9.7% of the sample reported exposure to IPV in the last 12 months. Correlations among the study variables are shown in Table 2. At the area level, womens education, living standards and urban residence are associated negatively with acceptance of mistreatment. Cross-level associations between analogous variables (e.g., womens education assessed individually and aggregated to the area level, r 0.664) are strong. Womens education and standards of living assessed individually exhibit the strongest associations with IPV (r 0.139 and r 0.173, respectively). Table 3 presents the multilevel null model of IPV in column one. IPV shows evidence of variability (clustering) at both the state (0.426) and community (0.481) levels equivalent to about 10.2 and 11.5 percent of total variance, respectively. In model 1, unadjusted associations between IPV and womens education assessed at the community and individual levels are statistically signicant, strong and negative in direction. The signicant quadratic term (OR 0.989) indicates that the reduction in risk of IPV is proportionally stronger at higher levels of womens education (non-linear effect). The random regression coefcients for the linear component of womens education are statistically signicant at the state (0.0013) and community levels (0.0046): these coefcients

  Logit pijk X 0ijk b Y 0ijk wjk nk wjk; 3ijk :

pijk is the probability of woman i, from community j, in state k, reporting IPV. X0 ijk is the vector of covariates (xed effects) corresponding to measures taken at the woman and community levels. Y0 ijk is a subset of covariates selected from X0 ijk at the woman level and allowed to vary randomly at the community level. nk, wjk, and 3;ijk are random effects intercepts unexplained residual variation at the state, community and woman-level, respectively. All estimates were derived using second order penalized quasilikelihood (PQL) and iterative generalized least-squares estimation in MLwiN (Rasbash, Steele, Browne, & Prosser, 2004). Residual variation at level 1 (3ijk) is assumed to have a standard logistic distribution with mean zero and variance p2/3 3.29. At subsequent levels (communities, wjk; states, nk), the ICC is given by the estimated residual variation at each level divided by total residual variation. In these models, the interpretation of xed effects estimates or the bs (e.g., the regression of IPV on womens education converted from logarithms of the odds ratios to odds ratios) is the same as it would be in ordinary logistic regression. Objective one is addressed by estimating IPV clustering at the state and community levels from the multilevel null model and then determining the unadjusted xed effects regression coefcients of IPV on womens levels of education assessed individually and aggregated to the community level. Objective two is addressed by allowing the regression of IPV on womens education to vary randomly at both the state and community levels (random regression coefcients: model 1). Objective three is addressed by controlling for womens attitudes towards mistreatment and living standards assessed at the community and individual levels. This is done in two steps to separate out the explanatory effect of womens attitudes towards mistreatment from their living standards. Objective four is addressed by including a cross-level interaction

Table 1 Sample characteristics. Level of analysis Variables Level 3, states (n 26) Number of communities M (SD) % prevalence IPV (min/max) Level 2, communities (n 3118) Number of women M (SD) % prevalence IPV (min/max) Womens education M (SD) Standard of living M (SD) Acceptance of mistreatment M (SD) % urban residence Level 1, women (n 68,466) Education in years M (SD) Standard of living M (SD) Acceptance of mistreatment M (SD) % exposed to mistreatment not IPV Age in years M (SD) % nuclear family Total number of children M (SD) % working % exposed to IPV n sample size; M mean; SD standard deviation. M (SD) % 119.9 (73.2) 8.5 (1.5/19.0) 22.0 (6.9) 9.5 (0.0/61.5) 3.9 (3.2) 0.01 (0.68) 0.01 (0.60) 31.3 3.9 (4.7) 0.00 (1.0) 0.00 (1.0) 4.9 31.8 (8.3) 51.4 3.1 (2.1) 35.0 9.7

M.H. Boyle et al. / Social Science & Medicine 69 (2009) 691697 Table 2 Pearson correlations among variables. Community levela 1. 2. 3. 4. Womens education Standard of living Acceptance of mistreatment Urban
b

695

1 1.0 .777 .338 .619 .664 .694 .205 .156 .007 .231 .179 .002 .119

2 1.0 .383 .588 .519 .680 .233 .153 .028 .153 .199 .013 .135

10

12

13

1.0 .231 .229 .303 .594 .074 .005 .101 .236 .048 .088

1.0 .408 .615 .138 .090 .038 .124 .167 .004 .068 1.0 .656 .212 .020 .094 .336 .170 .006 .139

Women level 5. Education in years 6. Standard of living 7. Acceptance of mistreatment 8. Age in years 9. Nuclear family 10. Total number of children 11. Working 12. Exposed to mistreatment (not IPV) 13. Exposed to IPV
a b

1.0 .218 .162 .175 .154 .212 .032 .173

1.0 .045 .010 .094 .152 .049 .117

1.0 .161 .524 .075 .012 .054

1.0 .151 .057 .017 .053

1.0 .100 .008 .044

1.0 .024 .074

1.0 .122

N 3118 clusters. N 68,466 women.

quantify between-state and between-community differences in the strength of association between IPV and womens education assessed individually. Omitted are the random effects covariances between the intercept and linear components of womens education and the quadratic components of womens education: all of them are small in magnitude and statistically non-signicant. In model 2, acceptance of mistreatment exhibits positive associations with IPV at both the community (OR 1.225) and individual (OR 1.324) levels. The inclusion of this variable attenuates the association between IPV and womens education but only at the community-level from OR 0.939 to OR 0.964. There is also a statistically signicant, cross-level interaction between acceptance of mistreatment at the area level and womens education in years OR 1.035: the protective inuence of womens education is muted in areas more accepting of mistreatment. In model 3, living standards exhibit a strong negative association with IPV but only at the individual level: OR 0.598. The inclusion of this variable attenuates all of the associations between IPV and womens

education. For example, the community-level effect for womens education is reduced from OR 0.964 to OR 0.982 and rendered non signicant. The individual level effect is reduced from OR 0.9190.960. There are also small reductions in the random regression coefcients for womens education at the state and community levels. In model 4, the remaining covariates show strong associations with IPV. However, the inclusion of these variables has only modest attenuating effects on associations between IPV and womens education. The largest impact is on the coefcient for living standards which goes from OR 0.598 to OR 0.666. Discussion In this study, clustering of IPV was extensive, accounting for about 21.5% of its variability and evenly divided between large states and communities. This contrasts with other studies in which the clustering of health-related phenomena is much less evident in

Table 3 Multilevel regressions of IPV on womens education, standard of living and acceptance of mistreatment in communities and individuals. Null model Fixed effects coefcients Level 2, community Womens education Acceptance of mistreatment Standard of living Urban residence Level 1, women (68466) Education in years Education in years squared Acceptance of mistreatment Standard of living Mistreatment not IPV Age in years Nuclear family Total number of children Working Interaction: mistreatment (L2) by education in years (L1) Fixed effect intercept, prevalence 6.37 7.02 Random effects variances, s2 (se) Level 3 (state) intercept 0.426 (0.124)* 0.332 (0.100)* Regression: womens ed 0.0013 (0.0006)* Level 2 (community) intercept 0.481 (0.029)* 0.337 (0.027)* Regression: womens ed 0.0046 (0.0014)* L1 Level 1, L2 Level 2, *p < 0.05. Model 1 Odds Ratio [95% CI] Model 2 Odds Ratio [95% CI] Model 3 Odds Ratio [95% CI] Model 4 Odds Ratio [95% CI]

0.939 [0.922; 0.957]*

0.964 [0.946; 0.984]* 1.225 [1.108; 1.354]*

0.982 [0.953; 1.012] 1.211 [1.095; 1.339]* 1.140 [0.996; 1.306]

0.992 1.192 1.043 1.171 0.956 0.991 1.293 0.666 3.840 0.975 1.250 1.055 1.361

[0.961; [1.077; [0.908; [1.049;

1.024] 1.319]* 1.198] 1.307]*

0.917 [0.900; 0.934]* 0.989 [0.986; 0.991]*

0.919 [0.904; 0.935]* 0.990 [0.988; 0.993]* 1.324 [1.283; 1.366]*

0.960 0.991 1.308 0.598

[0.944; [0.988; [1.268; [0.571;

0.975]* 0.994]* 1.350]* 0.627]*

[0.942; 0.970]* [0.989; 0.994]* [1.252; 1.335]* [0.634; 0.699]* [3.491; 4.223]* [0.977; 0.980]* [1.177; 1.326]* [1/037; 1.073]* [1.275; 1.452]*

1.035 [1.016; 1.055]* 6.62 0.372 (0.108)* 0.0009 (0.0005)* 0.310 (0.027)* 0.0047 (0.0014)*

1.033 [1.014; 1.052]* 6.25 0.395 (0.115* 0.0006 (0.0004) 0.312 (0.026)* 0.0045 (0.0014)*

1.036 [1.017; 1.055]* 4.20 0.486 (0.140)* 0.0004 (0.0003) 0.315 (0.027)* 0.0042 (0.0014)*

696

M.H. Boyle et al. / Social Science & Medicine 69 (2009) 691697

larger administrative areas than smaller communities in the same country (Boyle & Willms, 1999). States in India are relatively autonomous and quite diverse in language, culture and economic standing. Across states and communities in India, there are likely two distinct sets of social control mechanisms behind IPV one set that operates at administrative levels and reects laws, sanctions and institutional processes set in motion by state authorities, and another set that operates at the community level and reects local norms and social processes. Reconciling these two broad inuences on IPV could present serious challenges. The dominant inuence of womens education in this study is to reduce risk of exposure to IPV. At the individual level, the association between womens education and IPV is non-linear: the protective inuence of womens education is proportionately stronger at higher levels. This association is subject to area modication: in some states and communities it is stronger; in others, weaker. The identication of random effects such as these provides the motivation to test hypotheses about specic contextual factors that might serve as moderating inuences. Womens attitudes towards mistreatment is one such factor: the addition of this cross-level interaction in addition to individual level covariates explained most of the random effect for education at the state level but a relatively small amount of the random effect at the community level. At the community level, the association between womens education and IPV was accounted for by womens attitudes towards mistreatment and standards of living the former being the only contextual inuence to retain a statistically signicant association with IPV. Womens attitudes towards mistreatment was intended to represent among women, their collective tolerance of mistreatment, and to some extent capture broader norms, values and attitudes towards gender equality in the community. Tolerance of violence towards women at the community level has been identied as an important mechanism underlying womens risk of IPV in India (Segal, 1999) so the association between IPV and acceptance of mistreatment aggregated to the area level in this study comes as no surprise. However, it was also found that acceptance of mistreatment at the area level mutes some of the IPV risk reduction associated with higher levels of individual education. This interaction, modest in size, highlights the importance of considering the effects of person-context t in the occurrence of IPV. If women are to experience the full advantage of education in reducing their risk of exposure to IPV, it will be important to change attitudes towards the acceptance of IPV at the community level. Although multilevel modeling is useful for disaggregating area versus person effects, it cannot overcome study design limitations. To date, the examination of area effects on IPV has relied on crosssectional studies which are weak instruments for inferring causal signicance. Estimates of area-level inuences can be due in part, or wholly, to the characteristics of residents, often referred to as compositional or social selection effects. The allocation of persons to places is not random: socio-economic, cultural and kinship ties will act as lters causing people with shared features to settle in the same places. The general approach to this challenge is to identify confounding variables characteristics and experiences of women that account for IPV and are associated with differential settlement patterns and to control for their effects in the analysis. This is challenging to do for three reasons. One, causal understanding of IPV is limited, making it difcult to specify confounding variables in the rst place. Two, pertinent information needed to measure confounding variables, assuming that they can be identied, is rarely available in large omnibus surveys. Three, area effects on IPV will invariably be indirect and distal to the person. The fact that selection mechanisms sort individuals into communities does not preclude subsequent experience in these communities exerting

over time strong inuences on interpersonal behavior. Indiscriminately controlling for individual variables could attribute valid area-level inuences to confounding when, in fact, they have set in motion person-level processes that increase risk for IPV. In addition to uncertainty about the causal signicance of study variables, there are two other limitations that deserve comment. One, the measure of IPV is very general. The survey questions were limited to wife beating and excluded sexual assault; they provided no information on either severity or the circumstances in which violence occurred. There is no information on the reliability and validity of these questions, and concern exists about underreporting (Ellsberg, Heise, Pena, Agurto, & Winkvist, 2001). Two, the geographic boundaries that dene communities in this study were created for sampling purposes; there was no attempt to create socially meaningful clusters. However, the large magnitude of clustering in these small areas does suggest that meaningful differences exist between them. Ackerson et al. (2008) using the NFHS-2 demonstrated strong associations between IPV and womens education assessed individually and at the community level. The present study extends their work by showing that: 1) the protective inuence of womens education and IPV is non-linear (i.e., proportionately stronger at higher levels of education); 2) there is area-to-area variation in the association between womens education and IPV (random regression coefcients); 3) community differences in standard of living and womens attitudes towards mistreatment likely mediate the protective effect of womens education at that level; and 4) accepting attitudes towards the mistreatment of women at the area level mutes the protective effect of womens education assessed individually. The provision of educational programs is a basic responsibility of national governments and is often a foundation strategy for improving population health, national prosperity and global competitiveness. This study illustrates yet another important benet associated with providing educational opportunities for women. Although living in areas exhibiting attitudes highly acceptant for IPV mutes some of the protective inuence of education, the protective effect of education against exposure to IPV is both strong and robust. Of course, womens education is not the only factor associated with IPV. Changing attitudes towards the acceptance of IPV and improving population living standards can be expected to exert a strong downward pressure on risk for IPV.

References
Ackerson, L. K., Kawachi, I., Barbeau, E. M., & Subramanian, S. V. (2008). Effects of individual and proximate educational context on intimate partner violence: a population-based study of women in India. American Journal of Public Health, 98, 507514. Bangdiwala, S. I., Ramiro, L., Sadowski, L. S., Bordin, I. A., Hunter, W., & Shankar, V. (2004). Intimate partner violence and the role of socioeconomic indicators in WorldSAFE communities in Chile, Egypt, India and the Philippines. Injury Control and Safety Promotion, 11(2), 101109. Blakely, T. A., & Woodward, A. J. (2000). Ecological effects in multilevel studies. Journal of Epidemiology and Community Health, 54, 367374. Bollen, K. A., Glanville, J. L., & Stecklov, G. (2001). Socioeconomic status and class in studies of fertility and health in developing countries. Annual Review of Sociology, 27, 153185. Boyle, M. H., & Willms, J. D. (1999). Place effects for areas dened by administrative boundaries. American Journal of Epidemiology, 149, 577584. Dyson, T., & Moore, M. (1983). On kinship structure, female autonomy, and demographic behavior in India. Population and Development Review, 9, 3560. Ellsberg, M. C. (2006). Violence against women: a global public health crisis. Scandinavian Journal of Public Health, 34, 14. Ellsberg, M., Heise, L., Pena, R., Agurto, S., & Winkvist, A. (2001). Researching domestic violence against women: methodological and ethical considerations. Studies in Family Planning, 32(1), 116. Filmer, D., & Pritchett, L. H. (1999). The effects of household wealth on educational attainment: evidence from 35 countries. Population and Development Review, 25, 85120.

M.H. Boyle et al. / Social Science & Medicine 69 (2009) 691697 Filmer, D., & Pritchett, L. H. (2001). Estimating wealth effects without expenditure datador tears: an application to educational enrollments in states of India. Demography, 38, 115132. Gage, A. J. (2005). Womens experience of intimate partner violence in Haiti. Social Science & Medicine, 61, 343364. Garcia-Moreno, C., Jansen, H. A. F. M., Ellsberg, M., Heise, L., & Watts, C. (2005). WHO multi-country study on womens health and domestic violence against women: Initial results on prevalence, health outcomes, and womens responses. Geneva: WHO. Golding, J. M. (1999). Intimate partner violence as a risk factor for mental disorders: a meta-analysis. Journal of Family Violence, 14(2), 99132. Haub, C., & Sharma, O. P. (2006). Indias population reality: reconciling change and tradition. Population Bulletin, 61(3), 119. Hindin, M. J., & Adair, L. S. (2002). Whos at risk? Factors associated with intimate partner violence in the Philippines. Social Science & Medicine, 55, 13851399. Houweling, T. J., Kunst, A. E., & Mackenbach, J. P. (2003). Measuring health inequality among children in developing countries: does the choice of the indicator of economic status matter? International Journal for Equity in Health, 2, 817. International Institute for Population Sciences (IIPS) and ORC Macro. (2000). National Family Health Survey (NFHS-2), 199899: India. Mumbai: IIPS. Jewkes, R. (2002). Violence against women III: causes and prevention. The Lancet, 359(20), 14231429. Koenig, M., Ahmed, S., Hossain, M. B., & Mozumber, A. B. M. (2003). Womens status and domestic violence in rural Bangladesh: individual- and community-level effects. Demography, 40, 269288.

697

Koenig, M. A., Stephenson, R., Ahmed, S., Jejeebhoy, S. J., & Campbell, J. (2006). Individual and contextual determinants of domestic violence in north India. American Journal of Public Health, 96(1), 132138. Kravdal, O. (2004). Child mortality in India: the community-level effect of education. Population Studies, 58, 177192. MacMillan, H. L., Wathen, C. N., Jamieson, E., Boyle, M. H., McNutt, L.-A., Worster, A., et al. (2006). A randomized trial of approaches to screening for intimate partner violence in health care settings. Journal of the American Medical Association, 296(5), 530536. McQuestion, M. J. (2003). Endogenous social effects on intimate partner violence in Columbia. Social Science Research, 32, 335345. Plichta, S. B. (2004). Intimate partner violence and physical health consequences: policy and practice implications. Journal of Interpersonal Violence, 19, 12961323. Rasbash, J., Steele, F., Browne, W., & Prosser, B. (2004). A users guide to MLwiN version 2.0. London: University of London, Institute of Education, Multilevel Models Project. Segal, U. A. (1999). Family violence: a focus on India. Aggression and Violent Behavior, 4, 213231. Sharma, B. R. (2005). Social etiology of violence against women in India. The Social Science Journal, 42, 375389. Sugarman, D., & Frankel, S. L. (1996). Patriarchal ideology and wife assault: a metaanalytic review. Journal of Family Violence, 11, 1340. World Health Organization. (2002). World report on violence and health. Geneva: World Health Organization.