Methods
Based on a review of the analytical literature on child marriage [
4,
14‐
20], we developed a conceptual framework to guide the analysis. The framework posits that country-, community-, household-, and individual-level factors can play crucial roles in determining whether households decide to marry off their daughters prior to the legal age of marriage. The framework also hypothesizes that household decisions regarding whether to marry daughters as children and school enrollment are made simultaneously, not sequentially, to allow for the possibility that “those who intend to marry later (for whatever reason) tend to stay in school longer and those who intend to marry early leave school earlier to do so” [
3]. Our endogenous treatment of these two types of decisions follows Mensch et al. (2006) and Bajracharya & Amin (2010) [
3,
14]. It also follows the spirit of Becker’s application of neo-classical household economic theory to the analysis of marriage [
21].
Household- and individual-level determinants of child marriage include household wealth and income, the educational attainment of parents and relatives, religion, ethnicity, the age of the child and, and for girls, whether menarche has occurred. In addition to factors identified as important by economists, the framework incorporates factors emphasized by researchers from other disciplines, such as sociology (social norms, social networks) and anthropology (kinship systems). Community-level factors encompass social norms, which are typically defined as prevailing expectations within the community of appropriate behavior, or attitudes of approval and disapproval, specifying what ought and ought not to be done. These social norms include norms regarding gender, including child-rearing practices for boys and girls, and socialization. Also important is the concept of agency. Who in the household is making marriage decisions for children, and what role does the household have above and beyond the influence of community factors? Within a given community, it is assumed that households vary in terms of their willingness to marry their daughters early as a risk mitigation strategy – to broaden kinship in order to cope with the possibility of future health and economic shocks. The mother or other female relatives in the household may also have varying degrees of empowerment and attitudes regarding marriage, which can also be important. All of these factors can potentially influence not only the household decision of whether to marry off their daughters as a child, but also education decisions.
The primary source of data for the study is the 2010 Serbia Multiple Indicator Cluster Survey (MICS), which included a nationally representative sample of women of reproductive age throughout the country and a separate representative of women living in Roma settlements. The rationale for conducting a survey of Roma settlements to assess the wellbeing of Roma women and children, rather than oversampling persons who identify themselves as Roma in the national survey, is that many Roma may not define themselves as Roma due to fear of discrimination. In order to overcome this problem, a separate MICS of households living in Roma settlements was conducted, where it was assumed, based on previous research and government sources, that the vast majority of persons living in these settlements are in fact Roma [
7]. For the general household survey, data were collected from 6,392 households and 5,797 women of reproductive age (15 to 49 years of age). For the Roma settlements survey, data were collected from 1,711 households and 2,118 women of reproductive age. Ethical approval was sought and granted by the Statistical Office of the Republic of Serbia, based on the Law on Personal Data Protection. Informed consent procedures were administered to all survey respondents. Further details about the methodology and data collection procedures can be found in Statistical Office of the Republic of Serbia (2011) [
7]. The MICS data sets are publically available upon request by UNICEF.
Using the data from both surveys, both bivariate and multivariate analyses were carried out. The bivariate analyses were restricted to women 20–24 years of age and used to investigate the associations between child marriage (as measured by two dichotomous indicators – whether the individual reported entering a formal marriage or informal union before age 18 and before age 15) and the following household- and individual level characteristics: urban/rural residence, region, religion, household wealth, and educational attainment. We also investigated the bivariate associations between child marriage and attitudes towards domestic violence. The MICS included five questions on attitudes towards domestic violence. Questions were asked on whether the woman believes it is justifiable for a husband to beat his wife if she a) goes out without telling her husband, b) neglects the children, c) argues with her husband, d) refuses sex with her husband, and e) burns the food. Unfortunately, women were not asked whether they have ever been victims of domestic violence.
Multivariate analysis was conducted to explore how individual-, household-, and community-level factors were associated with child marriage and school enrollment among girls age 15 to 17 years. This age group was chosen because it is the only age group in which we had information on both marital status and school enrollment. The specific type of model we estimated was a bivariate probit model in which the outcome variables of interest were current marital status and school enrollment. Two separate models were estimated: one for girls from the nationally representative sample, and another for girls from the sample of households living in Roma settlements. For each model, the two dependent variables were marital status, as measured by a dichotomous variable of whether the girl was currently married or in union, and school enrollment, as measured by a dichotomous variable of whether the girl attended school in the current year.
There are two advantages of using a simultaneous modeling approach, as compared to a sequential modeling approach that estimates the probability of child marriage as a function of educational attainment. First, if there is interdependence among marriage and education decisions, modeling one outcome as a function of another may lead to biased parameter estimates of the association between education and the prevalence of child marriage. Second, the cross-sectional nature of the MICS does not permit the investigation of the sequential process of household decision-making regarding marriage and school enrollment decisions.
The independent variables consist of the age of the girl, relative household wealth (based on wealth quintiles generated through principle components analysis of household assets), religion, and urban/rural status. In addition, we included the following community-level measures in the models:
-
Percent of sampled adult females in the community who have zero tolerance of wife beating. This was measured by the percent of women in the community who do not justify wife beating for all five reasons: if wife goes out without telling husband; if wife neglects the children; if wife argues with husband; if wife refuses to have sex with husband; if wife burns the food.
-
Percent of sampled adult females in the community with secondary or higher levels of education.
-
Percent of sampled adult females in the community who were first married before age 18.
The sampling cluster was used to define the community, and in generating the community-level measures, we excluded the reference woman from the calculations to ensure that her characteristics did not influence the values of the indicators.
Descriptive statistics for the variables included in the model are presented in Table
1.
Table 1
Descriptive statistics for variables included in Bivariate Probit models
Residence |
Urban | 64.89 | 53.64 |
Rural | 35.11 | 46.36 |
Religion |
Orthodox Christians | 56.58 | 87.53 |
Others | 43.42 | 12.47 |
Household wealth |
Poorest | 19.06 | 14.84 |
Poorer | 20.19 | 18.68 |
Middle class | 19.31 | 27.84 |
Richer | 22.07 | 20.19 |
Richest | 19.37 | 18.44 |
Mean community-level indicators among adult females |
Secondary or higher levels of education | 12.56 | 82.87 |
First married before age 18 | 62.77 | 9.66 |
Zero tolerance of wife-beating | 75.41 | 96.43 |
N | 221 | 254 |
In the next section, we first present the results on the prevalence of child marriage among females 20 to 24 years of age for both the general population and those living in Roma settlements. To explore whether the extensiveness of the practice has changed over time, we then report the prevalence among older birth cohorts of females. We then report the results on the spousal age gap among females 15 to 19 and 20 to 24 years of age and on attitudes towards domestic violence. Next, we show patterns of child marriage practice among Roma females by various indicators of socio-economic status. To conclude the section, we present the multivariate results of the determinants of child marriage among females 15 to 17 years of age for both the nationally representative sample and the Roma settlements sample.
Discussion
The purpose of this study was to assess the risk factors associated with the practice of child marriage among females living in Roma settlements in Serbia, and to make comparisons with females living among the general population. The study contributes to the existing literature on child marriage in a number of ways. First, it is one of the few empirical studies of child marriage in a European setting. Second, it is one of the few studies that investigated child marriage among the Roma, and, by using data collected through a special purpose survey of households living in Roma settlements, avoided the problem of individuals not being willing to disclose their ethnic affiliation due to fear of discrimination. Finally, the study investigated whether child marriage and school enrollment decisions were interdependent, which to our knowledge has not yet been empirically explored in the child marriage literature.
The results indicated that, on average, Roma girls in Serbia were at very high risk of being married as children. Child marriage among Roma women in Roma settlements was also substantially higher than among women in the poorest wealth quintile in the general population. The vast majority of currently married Roma girls had partners who were less than 5 years older than themselves. However, for currently married girls in the overall population, the opposite is true – the vast majority had husbands more than five years older than themselves. The smaller spousal age gap among the Roma may be due in part to the fact that a greater percentage of Roma males are first married under age 18 than males in the general population [
7].
Similar to the findings of a large number of previous studies, child marriage was found to be associated with a number of socio-economic characteristics, including household wealth, education, and urban/rural status [
1,
3,
17]. While these findings suggest that economic factors may be playing an important role in influencing child marriage, it is important to emphasize that, for girls who married and moved away from their parent’s household to live with their husbands, we did not have information on the characteristics of the household where the girl grew up. It is possible that the relative socio-economic conditions of the individual’s present family and the family where she grew up may have been very different, particularly if girls tend to marry up the social ladder. We also explored whether child marriage was associated with community-level factors, as measured by the percentages of women in the community who: had secondary or higher levels of education; were first married under age 18; and who had zero tolerance of wife beating. These factors were found to be statistically significant among girls in the general population, but not among Roma girls, after controlling for other factors. The reason for the insignificance of these factors based on the Roma model estimations is unclear, but may be due to the low power of the analysis and the more limited variability in the community-level factors among this marginalized population. Urban/rural status was statistically insignificant after controlling for other factors in the model.
The tests for the endogeneity of current schooling and education based on the bivariate probit estimations provide some support for the hypothesis that decisions about girls’ current school attendance and child marriage are interdependent. For the nationally representative sample, that ρ is statistically significant (
p = 0.0) suggests that common unobserved factors influence decisions about girls’ current school attendance and child marriage. For the sample of females living in Roma settlements, there was only weak evidence of the influence of common unobserved factors on both outcomes
, as the statistical test on ρ is only marginally significant at the 10 % level (
p = 0.09, outside the conventional threshold of
p = 0.05). However, that rho is negative indicates that tradeoffs are being made between marriage and schooling in both groups, significantly more so among the general population, and that, in the Serbian context, it is not appropriate to disregard the possibility that school decisions are influenced by the timing of marriage. These findings are consistent with the arguments made by previous researchers regarding the interdependent nature of marriage timing and school enrollment decisions, due to the possibility that many of the same unobserved factors that influence the timing of marriage also influence schooling decisions [
3,
14]. The factors could include parents’ educational attainment and aspirations for children, perceptions regarding the returns to schooling, availability of employment opportunities, and the degree of bargaining power of the girl [
22].
The study has a number of limitations. First, the age and various outcomes investigated are self-reported and hence may be prone to bias due to social desirability and recall. Age of the respondent and age at marriage may be misreported. Second, for the bivariate probit models of the determinants of current marital status and current enrollment in school, the sample size of girls 15 to 17 years of age is relatively small, which reduced the power of the analysis. Third, the cross-sectional nature of the data prevented us from determining the actual temporal ordering of marriage and school enrollment decisions, nor did we have information on the types and characteristics of household members that played roles in the decisions. Fourth, the analysis of the determinants of child marriage and school enrollment includes a number of community-level indicators that were generated based on individual-level data from each sample cluster in the MICS data sets. While these factors may suggest the presence of a social norm, they do not necessarily establish the presence of a social norm. Furthermore, individuals’ perceptions of the boundaries of their community may not necessarily coincide with the boundaries of the sampling cluster. Individuals living in the same sampling cluster may not necessarily experience the same set of area influences especially in urban areas and outside the Roma settlements.
Competing interests
The authors declare that they have no competing interests.
Authors’ contributions
DRH, DG, AJG, and CC jointly conceptualized the study. DRH, DG, and AJG developed the methodological approach and DH and DG performed the statistical analysis. DRH and DG drafted the manuscript. All authors read and approved the final manuscript.