![Sample our Economics, Finance,Business & Industry journals, sign in here to start your access, latest two full volumes FREE to you for 14 days]({"type":"image" , "src" : "/sda/5931/TOC-Subject-Sample-2021-Economics-Finance-Business-Industry.jpg"})
Abstract
This study examines the causal effect of women’s age at marriage on prevalence of domestic violence using newly available household data from India. The paper employs an empirical strategy that utilizes variation in age at menarche to obtain exogenous variation in women’s age at marriage. The results show robust evidence that a one-year delay in women’s marriage causes a significant decline in physical violence, although it has no impact on sexual or emotional violence. Further, the study provides suggestive evidence that the effect of women’s marital age on physical violence arises because older brides, as compared to younger brides, are more educated and are married to more educated men. Overall, the findings underscore the importance of better enforcement of existing social policies that seek to delay marriages of women, as well as formulation of newer interventions, to reduce the prevalence of domestic violence in developing countries.
HIGHLIGHTS
The study examines the causal effect of marital age on exposure to domestic violence.
It utilizes recent household data from India.
Variation in age at menarche is used to obtain exogenous variation in age at marriage.
Results show one-year delay in women's marriage causes a decline in physical violence.
The study conducts further analysis to shed light on underlying mechanisms.
ACKNOLWEDGMENTS
The authors are very grateful to the editor, associate editor, three anonymous reviewers, Daniel Millimet, and session participants at the NEUDC Conference 2018 held at Cornell University, New York, the 14th Annual Conference on Economic Growth and Development held at the Indian Statistical Institute Delhi, and the 2nd International Conference on South Asian Economic Development at the South Asian University, New Delhi for comments and suggestions. All remaining errors are our own.
SUPPLEMENTAL DATA
Supplemental data for this article can be accessed at https://doi.org/10.1080/13545701.2021.1910721.
Notes
1 Although technically spousal violence or IPV is a subset of domestic violence, we shall use the terms domestic violence and IPV interchangeably throughout the paper since three-quarters of violence against women is intimate (Aizer Citation2010).
2 The NFHS 2015–16 data show that almost 30 percent women get married within three years of menarche and 60 percent within five years of menarche.
3 Our results are subject to a caveat. As we discuss, the data on age at menarche are available for women ages 15–25 in the survey. Thus, we had to restrict our analysis to women of this age group only.
5 The DHS surveys for all countries are available at https://dhsprogram.com/.
6 Some households in the State module did not have eligible women who could answer questions on domestic violence. Also, in some households in the State module domestic violence questionnaire could not be administered since privacy could not be obtained or due to other issues.
8 In households with more than one eligible woman, the woman administered the module was randomly selected through a specially designed sample selection procedure based on the Kish Grid, which was built into the household questionnaire.
9 For more on specificities about collection of data on domestic violence in NFHS, see NFHS data documentation (p. 496) available at http://rchiips.org/nfhs/NFHS-3%20Data/VOL-1/Chapter%2015%20%20Domestic%20Violence%20(468K).pdf. Also see NFHS surveyor training manual (p. 8) at http://rchiips.org/nfhs/Manuals/DV_Training_Manual.pdf.
10 We thank a referee and the associate editor of this journal for highlighting this issue.
11 See Supplemental Online Appendix A for additional details about the analytical sample.
12 The proportion of women who have faced at least one of the four kinds of domestic violence in our sample is 25 percent.
13 Note that both examples suggest that OLS estimates are likely to be biased downwards. In the first example, E[MarriageAge ε [X] < 0 and the coefficient of (unobserved) patriarchy would be positive. In the second example, E[MarriageAge ε[X] > 0, and the coefficient of (unobserved) ability would be negative. In principle, there might be other potential omitted variables which are not orthogonal to age of marriage of the women and might be correlated with their exposure to domestic violence.
14 Studies of twins have found that random genetic variation is the single largest source of variations in menarche (see, for example, Kaprio et al. [Citation1995]).
15 Later we use an alternative non-linear method of estimation to assess the robustness of our baseline results. However, for our baseline analysis we use a linear approach since, as noted by Wooldridge (2010), this procedure [IV-TSLS] is relatively straightforward and might provide a good estimate of the average effect. Joshua D. Angrist and Jörn-Steffen Pischke (Citation2008: 107) also argue:
while a nonlinear model may fit the CEF (conditional expectation function) for LDVs (limited dependent variable models) more closely than a linear model, when it comes to marginal effects, this probably matters little. This optimistic conclusion is not a theorem [but] … it seems to be fairly robustly true.
16 The early menarche group consists of those women who attained menarche at the age of 14 or earlier. The late menarche group consists of those women who attained menarche after the age of 14.
17 In Figure A1 in the Online Appendix, we also graph the scattered plot of average age at marriage by age at menarche with a linear fitted line. The graph shows clear evidence of a significant positive relationship between age at menarche and age at marriage.
18 Note, although not caste, spouse’s wealth level may be endogenous to marriage. For instance, parents who are in a hurry to marry their daughters might have a lower reservation quality of spouse, as reffected in their wealth. However, this is unlikely to cause the IV estimate of the effect of women’s age at marriage on domestic violence inconsistent since age at menarche is unlikely to be correlated with spouse’s wealth.
19 In Figure A2 in the Online Appendix, we also graph the scattered plot of average years of schooling by age at menarche with a linear fitted line excluding the outliers. The graph does not show any evidence of significant positive relationship between age at menarche and years of schooling.
Note, Sekhri and Debnath (Citation2014) and Chari et al. (Citation2017) also implicitly assume that age of menarche is not correlated with women’s education. Both the papers investigate the impact of marital age of the mother on child health and education outcomes. Marital age is instrumented by menarcheal age, but mother’s education is not controlled for. Given that mother’s education is conjectured to be a determinant of child outcomes, mother’s education becomes the part of the error term in the second stage regressions, which must be assumed to be uncorrelated to menarcheal age, for their second stage parameter estimates to be consistent.
20 The magnitude of the coefficients of age at marriage in the regressions with less severe and severe physical violence as outcome variables are quite large relative to the population average. One can argue that this might be due to how we have defined our outcome variables. To check whether the size of the estimated coefficients are sensitive to our baseline definition of outcome variables, in the Online Appendix, we follow Jeffrey R. Kling, Jeffrey B. Liebman, and Lawrence F. Katz’s approach (Citation2007) and create z-score based indices of domestic violence and repeat our baseline analysis. We continue to find large and economically significant effect of age at marriage on physical violence. This suggests that size of the estimated coefficients is not driven by our definition of the outcome variables.
21 It is worth noting that the IV estimates of age at marriage, in general, are larger than the corresponding OLS estimates. This might be because of omitted factors like classical patriarchy or ability of women. As discussed previously, if the omitted factor is classical patriarchy, the covariance between the omitted factor and marriage age would be negative and the coefficient of unobserved patriarchy should be positive implying the sign of the bias to be negative. For the case of omitted ability, the covariance is likely to be positive and the coefficient of unobserved ability should be negative again rendering the sign of the bias as negative. IV estimates could be larger than OLS estimates might be due to measurement error in age at marriage as well. Measurement error in marriage will tend to attenuate the OLS coefficients but not the IV ones. Further, as pointed out by Chari et al. (Citation2017), it is also important to note that the local average treatment effect interpretation of an instrumental variable estimate implies that we are estimating the causal effect of marriage and for the subpopulation whose marriage timing is affected by the instrument, that is, menarche. It is possible that causal effects for this subpopulation are larger than those for the population as a whole which might be the reason why we find the coefficient estimates from the IV regressions to be larger than those from the OLS regressions.
22 That under-recognition is a potential issue at least with sexual violence has been also noted by Anita Raj et al. (Citation2010) and Mosfequr Rahman et al. (Citation2014). For physical violence, however, this problem does not arise. Physical violence is well-defined, and even younger brides can easily say whether they were physically abused by their husbands or not. Using a simple two variable IV regression model in which emotional/sexual violence is regressed on age at marriage, and age at marriage is instrumented by age at menarche, it can be shown that if emotional/sexual violence is underreported (due to under-recognition) and this under-reporting varies inversely with age at marriage, the IV estimate of the coefficient of age at marriage obtained using the underreported emotional/sexual violence data is higher than the IV estimate of the coefficient of age at marriage that would be obtained if one had access to emotional/sexual violence data without misreporting (true IV estimate). This suggests that if sexual/emotional is underreported, it is possible for one to find the IV estimate of age at marriage to be zero (or even positive), when in fact the true IV estimate is negative.
23 Table A12 in the Online Appendix present the definitions and summary statistics of all these additional outcome variables.
24 This, in other words, mean that if marriage is delayed by a year, an average woman continues going to school for that full year and drops out of school right after marriage.
Additional information
Notes on contributors
Punarjit Roychowdhury
Punarjit Roychowdhury is Assistant Professor in the Industrial Economics Division at the Nottingham University Business School, University of Nottingham, UK. He obtained his PhD in Economics from Southern Methodist University, USA. Primarily an applied microeconomist, Punarjit’s research interests lie in the fields of microeconometrics, labor economics, development economics, and behavioral economics. His research papers have appeared in well-known journals including Journal of Business & Economic Statistics, Oxford Economic Papers, Oxford Bulletin of Economics & Statistics, and World Development, among others.
Gaurav Dhamija
Gaurav Dhamija is Assistant Professor at the Indian Institute of Technology, Hyderabad. He obtained his PhD in economics from Shiv Nadar University, India. His research interests lie in the fields of development economics, health economics, and labor economics. His research papers have been published in the Journal of Development Studies and Journal of International Development.