Local Crime and Early Marriage: Evidence from India

Abstract

This paper analyses whether living in a locality with high crime against women affects the probability of early marriage—that is, marriage before the legal age of marriage of girls. Using a nationally representative longitudinal data set and tackling the potential endogeneity of local crime rates, we find that perceived crime against women in the locality significantly increases the likelihood of early marriage of girls, while there is no such effect on boys of comparable age group. We also find no such effect of gender-neutral crimes (such as theft and robbery) on the likelihood of early marriage of girls. Moreover, we find that the relationship holds only in conservative households where the purdah system is practised, and also in the northern region of India, where patriarchal culture and gender norms are stronger than in the southern region. A sensitivity analysis assessing the potential impact of unobservable confounders suggests that our estimates are unlikely to be affected by omitted variable bias.

Keywords:

• crime
• early marriage
• gender
• India

JEL CLASSIFICATION:

• J12
• J13
• J16

1. Introduction

Early marriage of adolescent girls remains prevalent in various part of the world, especially in South Asia and Africa, despite extensive efforts to curb it. In India, 27% of females aged 20–24 married before 18 (NFHS 2015–16).¹ Previous research explores both causes and consequences of early marriage, with a specific focus on South Asia. These studies document that factors like poverty, parental education, limited opportunities, and social norms drive early marriage among young women (Mathur, Greene, & Malhotra, 2003; Oleke, Blystad, Moland, Rekdal, & Heggenhougen, 2006; Palermo & Peterman, 2009; Walker, 2012). On the other hand, consequences of early marriage include adverse effect on women’s health, human capital, vulnerability, post-marital agency, and the well-being of offspring from early childbearing (Jensen & Thornton, 2003; Maria Pesando & Abufhele, 2018; Sekhri & Debnath, 2014).² Against this backdrop, our study investigates a hitherto unexplored determinant of early marriage: how perceived crime against women in the locality plays a role in the marriages of adolescent girls in India.

A growing body of research has examined the impact of crime against women on their workforce participation and human capital development (Bowen & Bowen, 1999; Ceballo, McLoyd, & Toyokawa, 2004; Chakraborty, Mukherjee, Rachapalli, & Saha, 2018; Schwartz & Gorman, 2003). We add to this literature on the impact of crime against women by exploring its association with early marriage outcomes, particularly for adolescent girls. Our hypothesis posits that parents in high-crime localities, especially where crimes against women are prevalent, are more likely to arrange early marriages for their daughters due to the societal costs associated with female victims of sexual harassment. This effect is not expected for sons due to the patrilocal residence system. Focusing on different types of perceived crimes—gender-neutral and gender-specific—we empirically investigate the relationship between early marriage and crime against women, an aspect that, to our knowledge, has not been explored in existing literature on women’s empowerment and crime against women.

We analyze nationally representative household-level panel data collected in 2005 and 2012. The analysis focuses on adolescent girls aged 12 to 16 in 2005. The dataset includes tracking information for individuals who migrated between the two survey rounds. We particularly take advantage of this tracking data to get information on women who were married between the two survey rounds and migrated with their husbands. In absence of exogenous variation in perceived crime, it is difficult to identify the causal effects in our context. Nevertheless, acknowledging that perceived crime-against-women in the locality could be endogenous, we use this measure from the baseline survey conducted in 2005. This approach tackles the concern of reverse causality by regressing the marriage outcomes on the lagged measure of crime-against-women. Furthermore, we inspect the sensitivity of our estimates with respect to potential omitted variable bias following a method developed by Altonji, Elder, and Taber (2005) and Oster (2019). This analysis shows that our estimates are unlikely to be affected by omitted variable bias. We further elaborate on the issue of causal identification in our empirical framework in section 4.4.

We find that the likelihoods of marriage and early marriage increase by 5.7 percent and 9.6 percent on average with a one standard deviation increase in perceived crime against women in the locality. Results remain consistent after controlling for other gender-neutral crimes in the locality, background characteristics, and state fixed effects. We also investigate the potential mechanisms that drive the relationship between perceived crime against women and early marriage of adolescent girls. One channel that we test empirically is the value that conservative societies place on women’s chastity. We hypothesize that conservative families may marry off daughters early in response to perceived gender-specific threats. Results show a significant impact of crime against women on marriage outcomes only in households with conservative practices like purdah—the practice of screening women from men or strangers by covering their faces—and in northern states with stronger gender norms and patriarchal values.

Our study contributes to the existing literature on early marriage in many ways. First, it adds to the literature on the determinants of age at marriage by analysing an important but hitherto unexplored factor – crime against women in the locality. Second, the investigation is particularly important for India, where incidences of both early marriage of women and sexual violence are high. Crimes against women rose by 34 per cent between 2012 and 2015 (National Crime Records Bureau, 2016), and according to the UN India Business Forum (2018), 92 per cent of women in Delhi reported sexual or physical violence in public spaces. Moreover, in the conservative Indian society, the stigma surrounding young women who have been victims of sexual violence is particularly severe, as women’s chastity is strongly valued. In this context, it is important to investigate whether crime against women in the locality as perceived by the households disproportionately hurts adolescent girls more than boys of comparable age in India. This leads us to the third contribution of the paper: adding to the existing literature on gender inequality in India. Gender inequality in India exists in various forms—starting from sex imbalance at birth due to female feticide, unequal survival rates, inequality in health and educational expenditures, labour market discrimination, and so on. While the country has seen significant progress in reducing the gender gap in school enrolment, the quality of education and labour market outcomes of women have not improved commensurately (Klasen & Pieters, 2015; Sahoo & Klasen, 2021). Early marriage is negatively associated with human capital formation of girls as girls getting married early discontinue education (Delprato, Akyeampong, Sabates, & Hernandez-Fernandez, 2015; Sekine & Hodgkin, 2017). Low human capital formation in turn has the potential to lead to poor labour market outcomes. Thus, the findings of this study are also relevant to the discourse on gender inequality in India.

The rest of the paper is structured as follows. Section 2 discusses the existing literature on early marriage and effect of crime on women’s economic outcomes Section 3 outlines the research question and empirical method. Section 4 describes our data and details the construction of the main variables used in the analysis. Section 5 discusses the results and Section 6 concludes.

2. Literature review

The existing literature on early marriage has focused on both its drivers and consequences. Despite the adverse welfare consequences of child marriage being well established, the phenomenon is still pervasive in developing countries.³ Although most countries have a legal minimum age of marriage, in practice age of marriage in developing countries is determined by social norms. Adverse consequences of early marriage, as documented in the literature, are pervasive for the young brides. These adverse consequences include poor well-being in terms of mental and physical health, poor human capital formation, vulnerability, and lack of post-marital agency (Jensen & Thornton, 2003; Maria Pesando & Abufhele, 2018; Senderowitz, 1995). Various studies have shown that child marriage that subsequently leads to early childbearing has detrimental effects on the offspring (Sekhri & Debnath, 2014).⁴

Considerable research has identified a number of root causes or key drivers of child marriage. However, these drivers are often context-specific and depend on the country- or region-specific characteristics and institutions. Many studies have established an association between household poverty and girl child marriage (Dahl, 2010; Handa et al., 2015; Mathur et al., 2003). In societies with a patrilocal residence system, parents view daughters as responsibilities while sons are viewed as old-age security. Therefore, marrying off daughters relieves the parents of an economic responsibility. Moreover, the cost associated with marriage, called dowry, increases with the girl’s age, further pressurizing parents to marry their daughters off early. Lack of economic opportunities for women coupled with traditional gender roles has also been established as a driver of early marriage (Arends-Kuenning & Amin, 2000; Mathur et al., 2003). In an attempt to investigate the impact of role model and change in aspirations, Castilla (2018) estimates the effect of Panchayati Raj institutions that reserve seats at the district level for women Pradhans (or heads) on a rotating basis, on child marriage of girls in India. Results from the study indicate that exposure to women in local government decreases the likelihood of child marriage, and delay the age at first marriage. The author links the results to the literature of female role model and its impact on changing gender norms and aspirations of girls as well as their parents for the potential mechanism (Castilla, 2018).

Some of these studies have also highlighted the issue of purity concerns of young women once they reach puberty. As discussed by Mathur et al. (2003), once a girl reaches menarche the fear of potential pre-marital sexual activity and pregnancy becomes a major concern among family members who are accountable for ‘protecting’ her chastity and virginity until her marriage. This fear may lead to the decision to marry the girl off early to preclude any such ‘improper’ sexual activity. The safety and purity concerns of young women are naturally heightened if there are increasing incidences of crimes against women in the locality. However, none of the studies in the extant literature on early marriage has explored this issue.

In a recent article on female labour force participation, Chakraborty et al. (2018) view low female labour force participation in India as a response to fear of crime against women. Using the same data from the India Human Development Survey (IHDS), they show that women’s declining workforce participation in India can partially be accounted for by rising crime against women in the locality. In a similar line of work Siddique (2020) investigates the effect of media reports of violence on urban women’s labour supply decisions and find a temporary reduction in women’s labour supply outside home following increased media reports of sexual assaults. Mishra, Mishra, and Parasnis (2021) empirically investigate the effect of crime on women’s labour force participation and its gendered impact at the district level. The study finds that men’s labour force participation increased or remained unimpacted by an increase in crime, while an increase in crime decreased women’s labour force participation outside their home. While this recent literature has focused on crime against women in India and its impact on adult women’s decision to work, the impact of crime against women on marriage market outcomes such as likelihood of early marriage for adolescent girls is under researched. It ignores the fact that the effect of crime could start even earlier by forcing young girls to discontinue their education and get married at an earlier age compared to their male counterparts.⁵

We contribute to this literature by providing evidence on the effect of crime against women as perceived by the households on the likelihood of early marriage of young women while controlling for other gender-neutral crimes including physical attacks. We also examine the potential mechanisms through which this relationship is established, by investigating social norms and the stigma attached to being the victim of sexual harassment.

The findings from this paper have policy implications for reducing early marriage in developing countries. Despite repeated efforts by national governments and international development agencies to discourage and end the practice of early marriage, it remains prevalent. Current efforts, like awareness programs and cash transfers to reduce dropout and delay the age of marriage of girls, have not yielded the desired results. Currently, there are at least 15 such schemes (e.g. Apni Beti Apna Dhan programme in Haryana, Kanyashree Prakalpa in West Bengal and Beti Bachao Beti Padhao in northern states) in operation in India. However, these programmes have not been able to eradicate early marriage. Therefore, looking at other factors such as local-level safety and changing the perceptions of households may be an alternative policy instrument.

3. Methodology

To investigate the relationship between perceived crime against women in the locality and the likelihood of getting married, we first construct the dependent variable MarriedBetweenRounds dummy. It denotes whether an individual has been married between the two survey rounds in 2005 and 2012: $$\mathit{\text{MarriedBetweenRounds}}=\{\begin{array}{l}1\mathit{\text{ if Marital status in}}2012=\mathit{\text{Married}} \\ 0\mathit{\text{ if Marital status in}}2012=\mathit{\text{Unmarried}} \end{array}$$

The second dependent variable, $\mathit{\text{EarlyMarriage}},$ is a dummy variable denoting whether the individual married below legal age or not (including those married above legal age and those unmarried). This regression can be run on those who have crossed the legal age, hence the outcome variable is not censored. We define the dependent variable $\mathit{\text{EarlyMarriage}}$ as:⁶ $$\mathit{\text{EarlyMarriage}}=\{\begin{array}{l}1\mathit{\text{ if AgeMarriage}}<\mathit{\text{LegalAge}}(18 \mathit{\text{for girls and}} 21 \mathit{\text{for boys}})\\ 0\mathit{\text{ if AgeMarriage}}\ge \mathit{\text{LegalAge or Unmarried}} \end{array}$$ We estimate the following equations mainly for females: (1) $$\mathit{\text{MarriedBetweenRoun}}{\mathit{\text{ds}}}_{\mathit{\text{ihvs}}}={\beta }_0+{\beta }_{1}L\mathit{\text{ocalCrim}}{\mathit{\text{eBeforeMarriage}}}_{\mathit{\text{vs}}}+{\beta }_2{X}_{\mathit{\text{ihvs}} }+{\beta }_3{Z}_{\mathit{\text{hvs}}}+{\beta }_4{L}_{\mathit{\text{vs}}}+{\theta }_{s }+{\epsilon }_{\mathit{\text{ihvs}}}$$(1) (2) $$\mathit{\text{EarlyMarriag}}{e}_{\mathit{\text{ihvs}}}={\gamma }_0+{\gamma }_{1}L\mathit{\text{ocalCrim}}{\mathit{\text{eBeforeMarriage}}}_{\mathit{\text{vs}}}+{\gamma }_2{X}_{\mathit{\text{ihvs}}}+{\gamma }_3{Z}_{\mathit{\text{hvs}}}+{\gamma }_4{L}_{\mathit{\text{vs}}}+{\theta }_{s }+{ϵ}_{\mathit{\text{ihvs}}}$$(2)

Where $\mathit{\text{MarriedBetweenRoun}}{\mathit{\text{ds}}}_{\mathit{\text{ihvs}}}$ and $\mathit{\text{EarlyMarriag}}{e}_{\mathit{\text{ihvs}}}$ are the dummies representing whether an individual (female) i in the household h, village/town (or primary sampling unit) v, and state s, has been married between the two survey rounds and before the age of 18, the legal age of marriage for girls in India. The marriage outcome is observed in the follow up survey. The main variable we are interested in is a measure of crime rate at village/town, v, at the baseline survey given by $\mathit{\text{LocalCrim}}{\mathit{\text{eBeforeMarriage}}}_{\mathit{\text{vs}}}.$ The vector X includes individual specific observable characteristics such as age, education, and relationship to household head. Household specific characteristics such as household income, asset, occupation, caste, religion, household’s membership to women’s groups, and media exposure of the women in the household are included in the vector Z. Households that are members of women’s groups such as Self-Help Group and Mahila Mandal, and have more media exposure can be better aware and informed about the legal age of marriage for girls and the adverse impacts of child marriage. Therefore, inclusion of these variables in the regression is important to capture their influence on the marriage decision. Some village/urban PSU specific factors such as the location (urban or rural) and the number of girls enrolled in school in the locality can also influence the decision of a household to marry off a girl at a young age. To capture this we include proportion of girls enrolled in school in a PSU, and the location of the PSU (urban or rural) in the vector L. Our sample is spread across 33 states in India which have different culture and traditions. The decision of early marriage of girls can also be influenced by these state specific factors. Therefore, our models contain state fixed effects (${\theta }_{s }$) to control for state specific heterogeneity.

In Equations (1)(1) $$\mathit{\text{MarriedBetweenRoun}}{\mathit{\text{ds}}}_{\mathit{\text{ihvs}}}={\beta }_0+{\beta }_{1}L\mathit{\text{ocalCrim}}{\mathit{\text{eBeforeMarriage}}}_{\mathit{\text{vs}}}+{\beta }_2{X}_{\mathit{\text{ihvs}} }+{\beta }_3{Z}_{\mathit{\text{hvs}}}+{\beta }_4{L}_{\mathit{\text{vs}}}+{\theta }_{s }+{\epsilon }_{\mathit{\text{ihvs}}}$$(1) and (2)(2) $$\mathit{\text{EarlyMarriag}}{e}_{\mathit{\text{ihvs}}}={\gamma }_0+{\gamma }_{1}L\mathit{\text{ocalCrim}}{\mathit{\text{eBeforeMarriage}}}_{\mathit{\text{vs}}}+{\gamma }_2{X}_{\mathit{\text{ihvs}}}+{\gamma }_3{Z}_{\mathit{\text{hvs}}}+{\gamma }_4{L}_{\mathit{\text{vs}}}+{\theta }_{s }+{ϵ}_{\mathit{\text{ihvs}}}$$(2) we sequentially add crime rates for different types of crime, gender-specific and gender-neutral crimes to see their impacts on the likelihood of marriage and early marriage. These two equations are estimated primarily for women and thus do not tell us whether the effect of crime level is significantly different for women compared to men. Therefore, in Equation (3)(3) $$\mathit{\text{MarriedBetweenRoun}}{\mathit{\text{ds}}}_{\mathit{\text{ihvs}}}={\alpha }_0+{\alpha }_{1}L\mathit{\text{ocalCrim}}{\mathit{\text{eBeforeMarriage}}}_{\mathit{\text{vs}}}+{\alpha }_2{\mathit{\text{Female}}}_{\mathit{\text{ihvs}}}+{\alpha }_{3}L\mathit{\text{ocalCrim}}{\mathit{\text{eBeforeMarriage}}}_{\mathit{\text{vs}}}\times {\mathit{\text{Female}}}_{\mathit{\text{ihvs}}}+{\alpha }_4{X}_{\mathit{\text{ihvs}}}+{\alpha }_5{Z}_{\mathit{\text{hvs}}}+{\alpha }_6{L}_{\mathit{\text{vs}}}+{\vartheta }_{s }+{\mu }_{\mathit{\text{ihvs}}}$$(3) we introduce an interaction term between crime before marriage and the female dummy and estimate the equation for the entire sample of men and women: (3) $$\mathit{\text{MarriedBetweenRoun}}{\mathit{\text{ds}}}_{\mathit{\text{ihvs}}}={\alpha }_0+{\alpha }_{1}L\mathit{\text{ocalCrim}}{\mathit{\text{eBeforeMarriage}}}_{\mathit{\text{vs}}}+{\alpha }_2{\mathit{\text{Female}}}_{\mathit{\text{ihvs}}}+{\alpha }_{3}L\mathit{\text{ocalCrim}}{\mathit{\text{eBeforeMarriage}}}_{\mathit{\text{vs}}}\times {\mathit{\text{Female}}}_{\mathit{\text{ihvs}}}+{\alpha }_4{X}_{\mathit{\text{ihvs}}}+{\alpha }_5{Z}_{\mathit{\text{hvs}}}+{\alpha }_6{L}_{\mathit{\text{vs}}}+{\vartheta }_{s }+{\mu }_{\mathit{\text{ihvs}}}$$(3)

We expect ${\alpha }_1$ to be insignificant and ${\alpha }_3,$ the coefficient of our main interest variable, to be positive and significant. We estimate Equation (3)(3) $$\mathit{\text{MarriedBetweenRoun}}{\mathit{\text{ds}}}_{\mathit{\text{ihvs}}}={\alpha }_0+{\alpha }_{1}L\mathit{\text{ocalCrim}}{\mathit{\text{eBeforeMarriage}}}_{\mathit{\text{vs}}}+{\alpha }_2{\mathit{\text{Female}}}_{\mathit{\text{ihvs}}}+{\alpha }_{3}L\mathit{\text{ocalCrim}}{\mathit{\text{eBeforeMarriage}}}_{\mathit{\text{vs}}}\times {\mathit{\text{Female}}}_{\mathit{\text{ihvs}}}+{\alpha }_4{X}_{\mathit{\text{ihvs}}}+{\alpha }_5{Z}_{\mathit{\text{hvs}}}+{\alpha }_6{L}_{\mathit{\text{vs}}}+{\vartheta }_{s }+{\mu }_{\mathit{\text{ihvs}}}$$(3) using a linear probability model.⁷

4. Data and descriptive statistics

We use data from the IHDS.⁸ The IHDS is a nationally representative survey of 41,554 households in 1,503 villages and 971 urban neighbourhoods across India. It is a panel survey—the first round was surveyed in 2004–05 and the second follow-up survey was carried out in 2011–12. For convenience, we will refer the baseline survey year as 2005 and the follow-up survey year as 2012. Most of the households (83 per cent) and around 85 per cent of individuals surveyed during the first round were resurveyed during the second wave. The data contains information on a rich set of individual and household-level characteristics. However, the IHDS does not provide information on age at marriage for every individual. Only a specially administered questionnaire for women has this information, which is not useful for our study since these women are married women and came to reside in the sample households in a particular neighbourhood (village or primary sampling unit) after marriage.⁹ Therefore, the marital age of these women is not expected to be affected by the perceived crime rate of their husbands’ localities. Similarly, women who were born in the survey households and got married between two rounds had left the survey household (to live in their husbands’ households due to the patrilocal residence system in India), and therefore were not present in round 2. Therefore, construction of our dependent variables, married dummy and early marriage dummy, is not straightforward.

4.1. Construction of dependent variable

We, focus on two dependent variables: (1) $\mathit{\text{MarriedBetweenRounds}}$ dummy, and (2) $\mathit{\text{EarlyMarriage}}$ (marriage before legal age). For the first one, we simply look at those who were unmarried in round 1 and observe their marital status in round 2 of the survey. Those who got married between the two survey rounds (2005 and 2012) are assigned a value of 1, and those who remained unmarried are assigned a value of 0. This outcome is observed easily for those who are present in the household during the second survey round. However, those who moved out of the main household or migrated are not included in the household roster of round 2. For these women, we use the tracking data and the information on migrated individuals. The IHDS team has tracked the individuals who moved out of the original households and migrated to a different place. Around 57 per cent of females surveyed in the first round migrated between the two survey rounds. The survey team was able to track 72 per cent of these migrated individuals (Appendix Table A1). The tracking data has information on their education, marital status, year of migration, reason for migration, current place, occupation, and more. We use this information to construct our $\mathit{\text{MarriedBetweenRounds}}$ dummy.

For the EarlyMarriage dummy, we primarily rely on tracking data since the survey lacks age-at-marriage information for all ever-married individuals, obtaining it only for a subset of eligible married women. Using year of migration and reason for migration as proxies, we construct the age_at_marriage variable. As married women typically migrate for marriage, the year of migration serves as a reliable proxy for the year of marriage, particularly for women reporting ‘marriage’ as the migration reason. This approach is effective in our sample of young women, where 93.5 percent of those married between surveys migrated, and 95 percent of them cited ‘marriage’ as their migration reason. However, due to patrilocality, we can’t use migration year as a proxy for age_at_marriage for married men, as their migration reasons are primarily ‘work’ and ‘study’. Thus, our analysis of early marriage is restricted to the female sample. The EarlyMarriage dummy is constructed by assigning a value of 1 to those married before the legal age of 18 and 0 to those married or remaining unmarried after 18. In this way we may underestimate the incidence of early marriage if some or all of the sample women actually got married a few years before the migration year. However, we argue that if we see any effect of crime on the probability of early marriage, the effect would be a lower bound of the true effect.

4.2. Construction of the main independent variable: crime in the locality

Our main independent variable of interest is the household’s perception of crime in the neighbourhood. The data provides information about the perception of each household about different types of crime in their locality, such as conflicts, thefts, attacks/threats, and, most importantly, harassment of girls. Specifically, it asks ‘How often are unmarried girls harassed in your village/neighbourhood?’. The response is a categorical variable that takes values of 0 for never, 1 for sometimes, and 2 for often. The question is specifically asked for unmarried girls, and therefore is perfect to use in our study as our main focus is unmarried individuals in the first survey round. We aggregate the household responses to the neighbourhood level to construct our measure of perception of crime against women as the proportion of households in the neighbourhood who perceive that girls are harassed (responses 1 and 2) in their neighbourhoods. It could be argued that the households with more unmarried women may experience or perceive higher crime against women. To avoid this problem, we take the average of each of these reported crimes for the neighbourhood except the household itself. For example, the crime rate for the ith household in jth village is estimated by taking the average of crime rates reported by all other households in the jth village except the ith household.

4.3. Final sample and descriptive statistics

We restrict our sample of females to those who were 12–16 years of age in the baseline survey (i.e. in 2005) and were not married. We look at how their probability of getting married and probability of early marriage during the period 2005 and 2012 are affected by the crime rates of 2005. In this way, we do not observe the outcome and explanatory variables at the same time point. We chose this age group as most adolescent girls enter menarche at this age and, thus, are marriageable (the average age of menarche in India is 12.8 years as cited in the literature, e.g. Ramraj, Subramanian, & Vijayakrishnan, 2021; Sanyal & Ray, 2008). The analysis sets the upper age limit for girls at 16, creating a two-year gap to improve accuracy in capturing the impact of baseline crime on early-marriage decisions taken at a later time point. To elaborate, for girls who were 17 at the baseline but not married during data collection, their proximity to turning 18 (the legal age of marriage) reduces the potential influence of perceived crime in the locality on their under-age marriage decisions.¹⁰ Moreover, in 2012 when the outcome (marriage) is observed the sampled women are 19–23 years old, crossing the legal age of marriage (18).¹¹ Similarly, we restrict our male sample to a comparable age group, aged 15–19 in 2005 so that they are above the legal age of marriage (21) when the outcome is observed in 2012.

The sample size for unmarried women in the 12–16 years age group in 2005 is 12,392, with an average age of 13.86 and six years of average completed schooling in 2005 (Table A2). The outcome variables, marital status and early marriage, are observed in the follow-up survey in 2012. Information on marital status is obtained for 10,396 (84 per cent) of sampled women and the early marriage dummy is created for 9,963 women (80 per cent of the sample). This may lead to sample selection bias, as those who have missing information for the outcome variables could be systematically different from the others. We deal with this issue of sample selection in the robustness section. Around half (47.1 per cent) of the young women in our sample got married between 2005 and 2012. The rate of early marriage is 14.7 per cent for the sampled women with an average marriage age of 18, while the rate is 9.6 per cent for the sampled men of the comparable age group.¹² In terms of crime rates, 12 per cent of households report harassment of unmarried girls in the locality during the baseline survey.

We compare the estimates of early marriage and gender-specific crime rates from the IHDS data with other nationally representative survey data. For early marriage, we use the NFHS data and compare it with the estimates from the IHDS data in a scatterplot (Figure 1). The figure shows a positive relationship with a correlation coefficient of 0.5 that is also statistically significant. We also compare the gender-specific crime rates estimated from the IHDS data with the equivalent estimates from the National Crime Record Bureau (NCRB) data at the district level. The scatterplot between the estimates from these two sources is presented in Figure 2. The curve shows a weak negative or no relationship between the two estimates. This could be due to the differences in the data: the NCRB data captures the actual reporting of the crime rates at police station (and suffers from under-reporting due to various reasons such as stigma), while the IHDS data captures the households’ perception of crime against unmarried women in the locality. Therefore, these two estimates may not be correlated.

Figure 1. Scatter plot of early marriage rates, 2012.

Notes: The figure is a scatterplot for early marriage rates of girls as estimated from National Family and Health Survey (NFHS) and India Human Development Survey (IHDS) data for the year 2012. The early marriage rate for girls is defined as the percentage of girls in the 19–23 years age group in the year 2012 who got married before the age of 18.

Source: Author’s compilation based on data from the NFHS (wave 4) and IHDS (2011–12).

Figure 2. Scatterplot of gender crime rates, 2005.

Notes: The figure is a scatterplot between the gender crime rates obtained from National Crime Report Beauro (NCRB) data and India Human Development Survey (IHDS) data. Note that the NCRB data gives the crime rate calculated on the basis of actual reported crimes in a district, while the IHDS data gives households’ perception about crime in a district.

Source: Author’s compilation based on data from the NCRB and IHDS, 2004–05.

4.4. Discussion on identification

There are three potential challenges in identifying the causal effect of perceived crime in the locality on the marriage decision of adolescent girls and boys. The first challenge arises from the possibility of reverse causality. It can be argued that higher rate of early marriage of adolescent girls in a locality might result in fewer unmarried girls left in the locality, making these girls more likely to experience sexual harassment. In that case, it may be early marriage that leads to gender specific crime than the other way around. We address this problem by using the baseline measure of crime while the outcome is measured between the two rounds of survey. Specifically, we utilize the follow-up survey of 2012 to construct marriage outcomes that occurred between 2005 and 2012. Our explanatory variable is the perceived crime reported in the baseline survey conducted in 2005, and hence it is pre-determined with respect to the outcome variable. Thus, our empirical strategy addresses the issue of reverse causality.

Second, endogeneity may arise if our measure of perceived crime is self-reported by the household. For instance, households with unmarried girls might perceive greater threat of sexual harassment in the locality and, thus, might report higher level of gender specific crime in the locality. We mitigate this problem by taking an aggregate measure of the crime variable at the neighbourhood level excluding the household’s own perception. We have explained the construction of each of the crime measures in detail in the previous section (section 4.2).

Third, one can still argue that the estimate may suffer from omitted variable bias if the model does not account for all confounding factors affecting perceived crime in the locality as well as households’ decision to marry their daughters off. We include state fixed effects to control for regional-level unobserved heterogeneity; this would account for variation in gender norms across the states. Furthermore, we investigate the sensitivity of the estimated effects of perceived crime on marriage outcomes with respect to potential omitted variable bias following a method developed by Altonji et al. (2005) and Oster (2019). This analysis, described in section 5.4.4, shows that our estimates are unlikely to be driven by omitted variables.

5. Results

5.1. Main results

In Table 1 we present the result estimated from Equations (1)(1) $$\mathit{\text{MarriedBetweenRoun}}{\mathit{\text{ds}}}_{\mathit{\text{ihvs}}}={\beta }_0+{\beta }_{1}L\mathit{\text{ocalCrim}}{\mathit{\text{eBeforeMarriage}}}_{\mathit{\text{vs}}}+{\beta }_2{X}_{\mathit{\text{ihvs}} }+{\beta }_3{Z}_{\mathit{\text{hvs}}}+{\beta }_4{L}_{\mathit{\text{vs}}}+{\theta }_{s }+{\epsilon }_{\mathit{\text{ihvs}}}$$(1) and (2)(2) $$\mathit{\text{EarlyMarriag}}{e}_{\mathit{\text{ihvs}}}={\gamma }_0+{\gamma }_{1}L\mathit{\text{ocalCrim}}{\mathit{\text{eBeforeMarriage}}}_{\mathit{\text{vs}}}+{\gamma }_2{X}_{\mathit{\text{ihvs}}}+{\gamma }_3{Z}_{\mathit{\text{hvs}}}+{\gamma }_4{L}_{\mathit{\text{vs}}}+{\theta }_{s }+{ϵ}_{\mathit{\text{ihvs}}}$$(2) using a linear probability model (LPM). We use the LPM as we are interested in marginal effects and our model includes interaction terms that are easier to interpret when estimated through the LPM rather than non-linear models.Table 1 reports the marginal effects for our main variables of interest—crime levels in the locality as perceived/experienced by the households. Gender-specific crime is defined by the perceived threat of harassment of unmarried girls by the households. After controlling for a range of background characteristics from the baseline survey, gender-specific crime is significantly associated with an increase in the chance of getting married and chance of early marriage (before the legal marriage-age) for girls in the age group 12-16 in the baseline. The likelihoods of marriage and early marriage increase by 12.9 percentage points and 6.8 percentage points as we move from a crime free neighbourhood to a neighbourhood with 100 percent crime rate. These magnitudes also imply that a one standard deviation increase in the perceived gender crime in the locality is associated with 5.7 percent and 9.6 percent increase in the average likelihood of marriage and early marriage of girls, respectively.¹³

Table 1. Regression result: perceived crime in the locality and marriage decision of women

	Married between surveys		Early marriage
	(1)	(2)	(3)	(4)
	Gender specific crime	All crime	Gender specific crime	All crime
Gender specific crime: harassment of unmarried girls (baseline)	0.129***	0.109***	0.068***	0.060**
	(0.037)	(0.039)	(0.024)	(0.024)
Theft in the locality (baseline)		0.036		0.070
		(0.080)		(0.064)
Breaking-in in any household in the locality (baseline)		0.136		0.110
		(0.202)		(0.135)
Threat or attack in the locality (baseline)		0.142*		−0.034
		(0.084)		(0.057)
Constant	−0.375	−0.394	−0.456	−0.453
	(0.508)	(0.509)	(0.394)	(0.394)
Observations	9,799	9,789	9,392	9,382
R-squared	0.251	0.252	0.105	0.105

Notes: Standard errors clustered at the Primary Survey Unit or PSU (i.e. village or town) level are in parantheses.

*** p < 0.01,

** p < 0.05,

* p < 0.1. The sample includes young women who were 12–16 years old in the first survey in 2005 and unmarried at that time. The outcome variables (married between surveys and early marriage) are observed in the second survey round in 2012. All models include state fixed effects. Control variables include individual, household, and village/PSU level characteristics from baseline survey.

Source: Author’s compilation based on IHDS data.

In the third specification, we investigate whether other gender-neutral crimes—theft, breaking-in, and threat/attack—are also associated with marriage probability of young women. The results show that it is only the gender-specific crime that has a significant positive association with the probability of marriage of young women, along with threat and attack. This is not surprising, since both harassment of unmarried girls and threat and physical attack are crimes likely to cause substantial damage to a woman’s modesty. Similarly, the probability of early marriage significantly increases with the increase in gender-specific crime in the locality, whereas the gender-neutral crime (theft/robbery, breaking-in, and threat/attack) in the locality as perceived by the households does not have any significant relationship with the probability of early marriage.

The regression equations control for a rich set of covariates and state fixed effects.¹⁴ To capture the media exposure of women of the households, we use the aggregate index of radio, newspaper, and television use by women. In addition, we also include variables to capture the household’s representativeness in various women’s groups, such as Mahila Mandals and self-help groups. These variables are potential factors that should have an impact on the marriage decisions of girls in the households. Thus, controlling for these factors is necessary to avoid omitted variable bias. Our results remain significant after controlling for these variables in addition to individual characteristics, household characteristics, and some village-/PSU-level factors. The individual-level characteristics include age, age squared, education, and marital status. Household characteristics include caste, religion, occupation, income, assets, adult male education, and duration in the locality. We also control for an urban dummy and the proportion of girls in the age group 6–16 years enrolled in the school at the village or PSU level in all the regression specifications. The results with all the control variables are presented in Appendix Table A2.

We also conduct an analysis on the outcome variable age at marriage for those who were married between the survey rounds. The results (in Table 2) show that gender specific crime is associated with a significant decrease in the age at marriage for girls who were married. The reduction in age at marriage is around 0.31 year or around 4 months with an increase in the gender specific crime rate in the locality from no crime to high crime.

Table 2. Regression result: perceived crime in the locality and age-at-marriage of young women

	Age at marriage (in years)
	(1)	(2)
	Gender specific crime	All crime
Gender specific crime: harassment of unmarried girls (baseline)	−0.308*	−0.332*
	(0.162)	(0.182)
Theft in the locality (baseline)		−0.323
		(0.334)
Breaking-in in any household in the locality (baseline)		−0.869
		(0.933)
Threat or attack in the locality (baseline)		0.755*
		(0.397)
Constant	8.866***	8.875***
	(2.862)	(2.871)
Observations	4,174	4,170
R-squared	0.322	0.323

Notes: Standard errors clustered at the Primary Survey Unit or PSU (i.e. village or town) level are in parantheses.

*** p < 0.01,

** p < 0.05,

* p < 0.1. The sample includes young women who were 12–16 years old in the first survey in 2005 and unmarried at that time. The outcome variables are observed in the second survey round in 2012. All models include state fixed effects. Controls include variables at the individual, household, and village/PSU level from baseline survey.

Source: Author’s compilation based on IHDS data.

However, this analysis only considers the adolescent girls who got married between the rounds and excludes those who remained unmarried at the endline survey. Also, the variable of interest, age at marriage, measures time until the occurrence of the event, marriage. Therefore, for those women who have not yet experienced the event (and remained unmarried), we do not have information on their age at marriage, resulting in right censoring of the data. Hence, we use survival analysis that is suitable for modelling such processes. Specifically, we use the Cox proportional hazard model following the literature (Jayaraman, Gebreselassie, & Chandrasekhar, 2009). The results show that harassment of girls in the locality is associated with higher likelihood of marriage in both the specifications, with and without controlling for gender-neutral crimes in the locality (Table 3).

Table 3. Hazard rate estimated using the Cox proportional hazard model for age at marriage

	Hazard rate of marriage
	(1)	(2)
	Gender specific crime	All crime
Gender specific crime: harassment of unmarried girls	1.199**	1.223**
	(0.092)	(0.104)
Theft in the locality		1.164
		(0.200)
Breaking-in in any household in the locality		1.277
		(0.634)
Threat or attack in the locality		0.701*
		(0.125)
Observations	9,392	9,382

Notes: Standard errors clustered at the Primary Survey Unit or PSU (i.e. village or town) level are in parantheses.

*** p < 0.01,

** p < 0.05,

* p < 0.1. The sample includes young women who were 12–16 years old in the first survey in 2005 and unmarried at that time. The outcome variables are observed in the second survey round in 2012. All models include state fixed effects. Controls include variables at the individual, household, and village/PSU level from baseline survey.

Source: Author’s compilation based on IHDS data.

5.2. Comparison with men

Next we compare these results with that of men by estimating Equation (3)(3) $$\mathit{\text{MarriedBetweenRoun}}{\mathit{\text{ds}}}_{\mathit{\text{ihvs}}}={\alpha }_0+{\alpha }_{1}L\mathit{\text{ocalCrim}}{\mathit{\text{eBeforeMarriage}}}_{\mathit{\text{vs}}}+{\alpha }_2{\mathit{\text{Female}}}_{\mathit{\text{ihvs}}}+{\alpha }_{3}L\mathit{\text{ocalCrim}}{\mathit{\text{eBeforeMarriage}}}_{\mathit{\text{vs}}}\times {\mathit{\text{Female}}}_{\mathit{\text{ihvs}}}+{\alpha }_4{X}_{\mathit{\text{ihvs}}}+{\alpha }_5{Z}_{\mathit{\text{hvs}}}+{\alpha }_6{L}_{\mathit{\text{vs}}}+{\vartheta }_{s }+{\mu }_{\mathit{\text{ihvs}}}$$(3) with all controls including gender neutral crimes. We run the regression with the overall sample comprising both men and women. The regression includes an interaction term between the variable crime against women and the female dummy to test whether crime against women has any differential effect on the likelihood of marriage of women compared to men of a comparable age group. We do not find any significant association between the perceived gender-specific crime in the locality and likelihood of getting married of young men of aged 22–26, as presented in Table 4. However, the interaction term is significant and positive, implying significant gender differences in the effect of crime against women on the likelihood of marriage of young women. The results indicate that gender-specific crime in the neighbourhood as perceived by the households has an association only with the likelihood of marriage of women. In the next section we talk about the channel through which households’ perceptions of crime against unmarried women influences their decisions to marry their daughters off at an early age.

Table 4. Regression result: male and female

	Married between surveys
	(1)
	All control
Gender-specific crime: harassment of unmarried girls (baseline)	0.006
	−0.034
Female dummy	0.303***
	−0.011
Interaction term: Female dummy # gender crime	0.108***
	(0.038)
Constant	−0.565***
	(0.192)
Observations	18475
R-squared	0.196

Notes: Standard errors clustered at the Primary Survey Unit or PSU (i.e. village or town) level are in parantheses.

*** p < 0.01,

** p < 0.05,

* p < 0.1. The sample includes women aged 12–16 and men aged 15–19 in the first survey in 2005 and unmarried at that time. The outcome variables are observed in the second survey round in 2012. All models include state fixed effects. Controls include gender-neutral crimes, individual characteristics, household-level and village/PSU-level factors from the baseline survey as controls.

Source: Author’s compilation based on IHDS data.

5.3. Exploring potential mechanisms

There could be several mechanisms driving the positive effect of crime on likelihood of early marriage of girls. Our results (presented in Table 1) show that only gender-related crime (crime specifically targeting unmarried girls) has a positive relationship with likelihood of marriage and early marriage of young women. Gender-neutral crimes (such as theft and robbery) in the neighbourhood have no significant association with women’s likelihood of getting married early. Moreover, we also find that crime against women has no significant association with the likelihood of marriage of young men of a comparable age group. One explanation for these results is that the stigma cost that the society attaches to a victim of sexual harassment is high for unmarried young women, particularly in a society in which female chastity is valued highly in the marriage market. As found by Buss (1989) in a cross-country study of mate preferences, Indian men put more weight on their spouse’s sexual purity at marriage than on physical appearance (Buss, 1989).¹⁵ Therefore, delaying the marriage of girls, particularly those living in a locality with higher gender crime can be detrimental for their marital prospects.

Another possible mechanism could be that the parents living in localities with higher gender crime are concerned about the well-being of their daughters. Due to this concern, parents may decide to marry their daughters off as early as possible to a different village/town than the one they are residing. In this section we investigate these channels through which the positive impact of crime on marriage outcomes may work.

5.3.1. Social norms and female chastity

Our results show that only perceived gender-related crime (crime specifically targeting unmarried girls) has a positive relationship with likelihood of marriage or early marriage of young women. Perceived gender-neutral crimes (such as theft and robbery) in the neighbourhood have no significant association with women’s likelihood of getting married early. Moreover, we also find that crime against women has no significant association with the likelihood of marriage of young men of a comparable age group. One explanation for these results is that the stigma cost that the society attaches to a victim of sexual harassment is high for unmarried young women, particularly in a society in which female chastity is valued highly in the marriage market. As found by Buss (1989) in a cross-country study of mate preferences, Indian men put more weight on their spouse’s sexual purity at marriage The stigma attached to being a victim of sexual harassment is higher in a more conservative society with stronger gender norms, and consequently the age of marriage for women is lowered. We test our hypothesis by conducting our analysis on two sub-samples: one comprising households in which the women practice purdah (the practice of screening women from men or strangers by covering the face), and another in which this practice is not present. Using the same Indian data, Desai and Andrist (2010) find that the practice of purdah or ghunghat, male–female segregation in the household, and restricted female mobility are all associated with early age at marriage. The survey asks each woman member of the households about practicing this ritual, purdah. We use this information to divide the sample into two groups. We expect to see a stronger positive impact of crime against women on the marriage outcomes of girls from the purdah practicing households. This expectation is based on the assumption that the value placed on female chastity will be higher in the households belonging to conservative group (or purdah sample).

We also test this hypothesis by dividing the sample into two regions: the northern region and the southern region. The existing literature shows that gender norms and patriarchal culture are stronger in the northern states of India than in the southern states (Dyson & Moore, 1983; Eswaran, Ramaswami, & Wadhwa, 2013; Sarkar, Sahoo, & Klasen, 2019). We hypothesize that the north–south difference may also manifest itself in the decision of parents to marry off their daughters at an early age if there is a perceived threat of gender-specific crime in the locality. Hence, we estimate the model separately for northern and southern states and test whether the effects of gender-specific crime on marriage/early marriage are different.

In Table 5 we present the results for both groups and for both the dependent variables $\mathit{\text{MarriedBetweenRouds}}$ dummy and $\mathit{\text{EarlyMarriage}}$ dummy. The results confirm our hypothesis. Crime against women has a significant positive association with likelihood of both marriage and early marriage only in the households where women practice purdah. Similarly, crime against women is significant and positive only in the sample for the northern region; the coefficients are insignificant in the sample for the southern region.

Table 5. Regression result: gender norms

	Married between surveys				Early marriage
	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
Variables	Purdah	No Purdah	North	South	Purdah	No Purdah	North	South
Gender specific crime: harassment of unmarried girls (baseline)	0.136***	0.029	0.158***	0.000	0.074**	0.015	0.105***	−0.030
	(0.049)	(0.055)	(0.046)	(0.079)	(0.034)	(0.034)	−0.03	(0.042)
Constant	−0.302	−0.148	−0.126	−1.012	−0.659	−0.022	−0.379	−0.355
	(0.692)	(0.841)	(0.620)	(1.030)	(0.570)	(0.590)	(0.497)	(0.796)
Observations	5,201	3,595	6,382	2,641	4,997	3,447	6,114	2,514
R-squared	0.269	0.255	0.253	0.254	0.125	0.102	0.113	0.098

Notes: Standard errors clustered at the Primary Survey Unit or PSU (i.e. village or town) level are in parantheses.

*** p < 0.01,

** p < 0.05,

* p < 0.1. The sample includes young women who were 12–16 years old in the first survey in 2005 and unmarried at that time. The outcome variables are observed in the second survey round in 2012. All models include state fixed effects. Controls include gender-neutral crimes, individual characteristics, and household-level and village/PSU-level factors from the baseline survey as controls. State fixed effects are included in all the regressions.

Source: Author’s compilation based on IHDS data.

5.3.2. Well-being concerns

While protecting female chastity may be one reason that the parents worry in locality with higher gender specific crime, it is also possible that the parents are concerned about the well-being of their daughters. If this is true then the parents may decide to marry their daughters off to a different village/town than the one they are residing. We test this hypothesis by using information on the current location of members not part of the baseline sample household or migrated for any reason including marriage. This information is available in the tracking data indicating whether a member not present in the baseline sample is living in 1) separate household within same building/compound, 2) separate household in the same village/urban locality, 3) somewhere else. Combining this information with the marriage outcome we generate a categorical variable indicating whether a girl was married outside the village/urban PSU or in the same village/urban PSU or remained unmarried. Next, we run a multinomial logit regression using this categorical variable as the outcome and gender specific crime rate as the main variable of interest. The results (marginal effects) from the multinomial logit regression presented in Table 6 show that gender crime significantly increases the likelihood of marriage outside the village/urban PSU comparing with the base outcome category ‘remained unmarried’. The marginal effect of gender specific crime is insignificant for the likelihood of marriage in the same village/urban PSU.¹⁶

Table 6. Regression result: marriage outside the village/PSU

	Marriage outside the village/PSU (Multinomial Logit regression)
	(1)	(2)
	Outside the village/PSU	Same village/PSU
Gender specific crime: harassment of unmarried girls (baseline)	0.151***	0.003
	(0.049)	(0.010)
Observations	9789	9789

Notes: Marginal effects are presented from a multinomial logit model. Standard errors clustered at the Primary Survey Unit or PSU (i.e. village or town) level are in parantheses.

*** p < 0.01,

** p < 0.05,

* p < 0.1. The sample includes young women who were 12–16 years old in the first survey in 2005 and unmarried at that time. The outcome variables (marriage outside village/PSU) are observed in the second survey round in 2012. All models include state fixed effects. Control variables include gender neutral crimes, individual characteristics, and household-level and village/PSU-level factors from the baseline survey.

Source: Author’s compilation based on IHDS data.

5.4. Robustness analysis

We present various robustness checks of our main results on marriage outcomes presented in Table 1 in this section. The results from the robustness analysis are presented in the sub-sections below.

5.4.1. Sample selection due to attrition

Our 2005 sample of adolescent girls experienced attrition in the 2012 follow-up survey, primarily due to migration. Some migrated members were tracked, and their information was collected, while others couldn’t be traced, resulting in missing data on marital status and early marriage for 16% and 19.6% of sampled women, respectively. These dropouts may introduce bias if non-random and from high- or low-crime areas compared to the analyzed sample. To assess this, we estimate an attrition equation, regressing attrition on crime rates and control variables (see Appendix Table A4). Results indicate that crime rates are not significant, suggesting that selection bias based on crime rates is unlikely. In summary, those not in the sample, do not seem to be from higher or lower crime areas than those included.¹⁷

5.4.2. Controlling for conservative attitudes of households

We attempt to mitigate the concern of endogeneity by including various household-level factors. To capture conservative attitudes, we incorporate a dummy variable indicating whether households practice purdah and have a male-first meal tradition. These practices are indicative of conservative views. Adding these variables to the regression helps control for household conservatism to some extent. The results in (Appendix Table A5) show smaller but still significant coefficients for crime against women in relation to both marriage likelihood and early marriage. As anticipated, the purdah and men-eat-first dummies are significant and positive, suggesting that households with these practices have a higher likelihood of early marriage for adolescent girls in India.

5.4.3. Considering village infrastructure and shocks for the rural sample

Beyond societal stigma, the influence of perceived crime on early marriage may hinge on village infrastructure and household shocks in rural areas. Factors like proximity to police stations, transportation, and educational institutions can shape perceptions of gender-specific crime and influence decisions on early marriage. Shocks such as droughts or floods may impact both crime perception and marriage decisions for girls. To account for these effects, we incorporate distance to key locations and information on village shocks in the regression, including squared terms for non-linear effects. Despite these additional village-level controls, the analysis in (Appendix Table A6) shows that the effect of crime against women persists. Notably, the impact on marriage likelihood and early marriage is higher in the rural sample, at 14 and 11.6 percentage points, respectively, compared to the overall sample.

5.4.4. Assessing potential bias from unobservables: Altonji-Oster method

While we demonstrate the robustness of our estimated effect by controlling for various confounding factors, potential omitted variable bias could arise from unobservable variables correlated with both the main explanatory variable and the outcome. To assess this bias, we apply the method developed by Altonji et al. (2005) and extended by Oster (2019), examining the impact of observables on the main coefficient, gender-specific crime. The analysis relies on two key parameters: δ representing the relative impact of unobservables compared to observables, and R²_max, the hypothetical regression’s R² controlling for all covariates (observables and unobservables). Oster (2019) shows that a bias adjusted coefficient can be estimated based on these two parameters. Analyzing the sensitivity of the coefficients from a set of published papers, she also suggests that a suitable value of R²_max is 1.3 times the R² of the regression that includes all the control variables (R²_controlled).¹⁸ For a reasonable value of δ, we follow the existing literature and assume δ = 1, implying that the observable control variables are as important as the unobservables (Altonji et al., 2005). Alternatively, we also show the estimate of δ that would make the main explanatory variable to have null effect on the outcome.

Results in (Appendix Table A7) show that the identified coefficient remains stable, indicating robustness. The first two columns display estimates from the regression: one without any control variables and the other including all control variables. Even under conservative assumptions, the effect of unobservable factors would need to be unreasonably high (16–18 times the effect of observables) to nullify the main coefficient. This exercise suggests that our main regression estimate, including all control variables, is stable and unlikely to be confounded by unobservable factors

6. Conclusion

We study the relationship between the likelihood of early marriage of men and women and perceived crime levels in the neighbourhood. Both gender-neutral crimes and gender specific crime are explored. Our results suggest that an increase in perceived crime against women in a locality significantly increases the likelihood of marriage and early marriage of women, while it does not affect the likelihood of marriage of men of comparable age group. The result holds after controlling for a range of individual- and household-specific factors and state fixed effects.

We also investigate the channels through which crime against women influences the decisions of households to get adolescent girls married off before the legal age for marriage. We argue that the positive association between perceived crime against women and their early marriage exists because of the concern regarding preservation of young women’s chastity, and the social stigma attached to victims of sexual violence, particularly in more conservative households. Our results support this argument by providing evidence that this positive relationship holds only in households where the purdah system—screening of women from men or the outside world by covering their faces—is practised, and also in the northern region of India, where patriarchal culture and gender norms are stronger than in the southern region as established in the literature (Dyson & Moore, 1983; Eswaran et al., 2013). In societies where female chastity is highly valued and rewarded, perhaps more than their education, parents are pressurized to marry off their daughters early, for fear of sexual harassment or sexual activity before marriage. Higher levels of crime against women in the locality, including rape, molestation, eve-teasing, etc., naturally heighten these fears, and lower the age of marriage.¹⁹

The phase between adolescence and adulthood is an important time in the lives of young people. The transition to marriage and subsequent fertility has serious implications for a woman’s future life trajectory and well-being, particularly in developing countries. Discontinuation of education after puberty and early marriage can contribute to the issue of low female labour force participation. Therefore, the age at which these transitions occur is a key concern as it is associated with human capital development, labour market outcomes, health, vulnerability, and the vicious circle of poverty.

Current policies addressing girl child marriage emphasize adherence to legal age requirements and the provision of incentives to promote school retention. Our study underscores the need for safety in localities to ensure a secure environment for girls. It highlights the discrepancy between reported crimes and the fear families may experience. Our findings advocate for policies addressing both actual crime and perceived fear of harassment of girls. This fear may stem from societal values on women’s chastity and a lack of trust in institutions.

Acknowledgements

The author would like to thank Elena Gross, Maria Lo Bue, Soham Sahoo, Kunal Sen, two anonymous referees and the Editor of this journal, and the participants of the UNU-WIDER workshop on women’s work for their valuable comments and discussions on the first draft of the paper. The author is also grateful to Dipanwita Ghatak for her excellent research assistance. The financial support for this research from the Economic and Social Research Council (grant ref no. ES/T010606/1), UK and the United Nations University World Institute for Development Economic Research (UNU-WIDER), Finland is gratefully acknowledged.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

This work was supported by the Economic and Social Research Council (ESRC), UK and The United Nations University World Institute for Development Economic Research (UNU-WIDER), Helsinki, Finland.

Notes

1 Despite global efforts by international organizations, governments, and non-governmental organizations (NGOs) to raise awareness about the adverse effects of early marriage and the implementation of the incentives for parents to delay their daughters’ marriage, the prevalence of early marriage among female adolescents remains high, particularly in India.

2 While early marriage is an issue for both genders, it has particular implications for females (Jensen & Thornton, 2003; Maria Pesando & Abufhele, 2018). Early marriage has been associated with withdrawal of adolescent girls from education and limited engagement with the labour market, as well as low literacy rates, increased risk of sexual violence, and poor health outcomes for women and their offspring (Bhanji & Punjani, 2014; Nour, 2009; Zahangir & Kamal, 2011).

3 Using Demographic and Health Survey data from 48 countries for the period 1986–2010, a United Nations study found little improvement in the practice of child marriage in both rural and urban areas (UNFPA., 2012). It is also important to note two other stylized facts about female early marriage practices. Historically, the practice has been widely prevalent in China, the Middle East, and the Indian sub-continent (Dixon, 1971), and absent from Europe from at least the beginning of the eighteenth century, when reliable records began (Hajnal, 1965). Second, the practice is most prevalent today in the least developed countries (UNICEF, 2018).

4 A recent literature investigates the impact of early marriage on women’s labour force participation (Assaad et al., 2020; Dhamija & Roychowdhury, 2020; Sunder, 2019). The findings are mixed. Assaad et al. (2020) and Sunder (2019) provide evidence to support the claim that early marriage reduces labour market participation of women in MENA region and Uganda. However, a recent study by Dhamija and Roychowdhury (2020) on India finds a delay in women’s age at marriage has no significant causal effect on their labour market outcomes.

5 In India most marriages are arranged by the parents. In a 2018 survey of more than 160,000 households, 93% of married Indians reported that their marriage was arranged by the family. Just 3% had a ‘love marriage’ and another 2% described theirs as a ‘love-cum-arranged marriage’, which usually indicates that the relationship was set up by the families, and then the couple agreed to get married.

6 The Prohibition of Child Marriage (Amendment) Bill, 2021, passed by the Lok Sabha, sought to amend the Prohibition of Child Marriage Act, 2006, to increase the minimum age of marriage for women from 18 to 21 years. However, this paper uses data from the year 2004-05 and 2011-12 when the legal minimum age of marriage for women was 18 years and for men was 21 years. Therefore, in this paper we use 18 years and 21 years age cut-off to define early marriage or marriage before legal age of marriage for women and men respectively.

7 Due to data limitations, we restrict the analysis of the overall sample including males to estimating the likelihood of marriage; analysis of early marriage is not conducted for the male sample.

8 The survey was carried out jointly by the University of Maryland and the National Council of Applied Economic Research, New Delhi. The dataset is publicly available at https://ihds.umd.edu.

9 The primary sampling unit (PSU) in the survey is a village for rural area and a town for urban area. Our main variable of interest, crime against women, is defined at the PSU level. We use the words neighbourhood or locality or village/PSU interchangeably in this article.

10 An analysis using 17 year old girls in the sample is presented in Table S2 in supplementary document. The results remain unchanged.

11 Note that all women in our sample have crossed the legal age of marriage – therefore the outcome of early marriage is fully observed for all of them.

12 The rate of early marriage among women aged 19–23 is 26 per cent in India as estimated from another nationally representative survey, the National Family and Health Survey (NFHS) in 2012.

13 The standard deviation (SD) of perceived crime against women is 0.207 as presented in Table A2. Therefore, 1 SD increase in crime against women leads to 2.7 (1.4) percentage point increase in the likelihood of marriage (early marriage). Considering the average likelihoods of marriage (0.471) and early marriage (0.147), these estimates translate into 5.7 percent and 9.6 percent increase in the respective average likelihoods.

14 It is also possible to use district fixed effects and control for the district specific heterogeneity that can influence marriage decision and perceived gender specific crime in a locality. We conduct the analysis using district fixed effects. We find that the results mostly hold for $\mathrm{\text{MarriedBetweenRouds}}$ dummy although there is a fall in magnitude and the level of significance. However, the coefficient of crime against women becomes insignificant in the $\mathit{\text{EarlyMarriage}}$ regression, possibly because there is not enough variation left in our main explanatory variable after including district fixed effects in the regression. Since India is a large country where gender norms have great variation across states but are relatively homogenous within states, we believe that inclusion of state fixed effects would adequately control for the unobserved factors such as gender norms. Therefore, we present our main results from regressions using state fixed effects and present the results from district fixed effects in an appendix Table A3.

15 The study also found men from China, Indonesia, Taiwan, and Iran revealed the same preference, while the opposite prioritization was seen in each of the 24 European, North American, South American, and sub-Saharan African countries included in the study (Buss, 1989).

16 Due to lack of variation in the location of marriage in the early_marriage sample we restrict this analysis only to the outcome, probability of marriage between rounds. More than 99% of girls who were married before legal age of marriage were married outside the village/urban PSU. Therefore, the analysis of early marriage outside village/PSU or in the same village/PSU is not possible.

17 Even though we are less worried about the sample selection as the attrition doesn’t seem to be systematically different from the retention sample based on crime rates, we use the Heckman selection correction model to check any selection bias arising from the sample attrition (Heckman, 1981). The results (not presented in the paper) show that crime against women remains significant after correcting for the selection bias.

18 Since R²_max cannot exceed 1, therefore it implies that R²_max = min{1.3*R²_controlled, 1}. We use 1.3*R²_controlled as it is always less than 1 in our case.

19 Eve-teasing is a form of sexual harassment practised generally by a man or a group of men to annoy women. Examples include verbal abuse by making sexual comments in public places.

Appendix A

Table A1. Sample retention in surveys

	Male		Female		All
	N	%	N	%	N	%
Present in both surveys	7,438	66.57	5,291	42.7	12,729	54.01
Migrated between surveys (2005 and 2012)	3,736	33.43	7,101	57.3	10,837	45.99
Tracked in 2012	2,045	54.74	5,136	72.33	7,181	66.26
Could not be tracked	1,691	45.26	1,965	27.67	3,656	33.74
Total	11,174	100	12,392	100	23,566	100

Notes: The sample includes unmarried females, 12–16 years old, and unmarried males, 15–19 years old in 2005.

Source: Author’s compilation based on IHDS data.

Table A2. Summary statistics

	Female			Male
Variable	Obs.	Mean	Std dev.	Obs.	Mean	Std dev.
Dependent variables
Married dummy	10,396	0.471	0.499	9,467	0.344	0.475
Early marriage dummy	9,963	0.147	0.354	8,160	0.096	0.294
Age at marriage (in years)	4,463	18.3	1.967	–	–	–
Main variables of interest : crime (baseline)^α
Gender-specific crime: harassment of unmarried girls in the locality (baseline)	12,311	0.126	0.207	11,090	0.123	0.203
Gender-neutral crimes (baseline)
Theft in the locality (baseline)	12,367	0.04	0.091	11,143	0.038	0.088
Breaking-in at any household in the locality (baseline)	12,369	0.010	0.033	11,145	0.010	0.033
Threat or attack in the locality (baseline)	12,369	0.028	0.078	11,145	0.026	0.072
Individual characteristics (baseline)
Age	12,392	13.867	1.431	11,174	16.854	1.378
Age square	12,392	194.353	39.918	11,174	285.957	46.583
Years of education completed	12,357	6.043	2.977	11,139	7.799	3.458
Parental years of education	11,955	3.085	4.299	10,708	2.905	4.204
Media exposure of women	12,174	0.456	0.332	10,954	0.467	0.333
Household characteristics (baseline)
Caste: Other caste	12,392	0.443	0.497	11,174	0.448	0.497
Caste: Other Backward Class	12,392	0.211	0.408	11,174	0.209	0.407
Caste: Scheduled Caste	12,392	0.088	0.284	11,174	0.084	0.277
Caste: Scheduled Tribes	12,392	0.258	0.438	11,174	0.259	0.438
Religion: Hindu	12,392	0.767	0.423	11,174	0.778	0.416
Religion: Muslim	12,392	0.156	0.363	11,174	0.146	0.353
Religion: Other	12,392	0.077	0.267	11,174	0.076	0.266
Relationship to head: daughter/son of head	12,392	0.848	0.359	11,174	0.872	0.334
Relationship to head: granddaughter/son	12,392	0.111	0.314	11,174	0.075	0.263
Relationship to head: other relationship	12,392	0.041	0.198	11,174	0.051	0.220
Years living in the locality	12,380	72.976	30.018	11,158	72.594	30.059
Household occupation: agriculture	12,392	0.263	0.440	11,174	0.265	0.441
Household occupation: casual labour	12,392	0.319	0.466	11,174	0.312	0.463
Household occupation: artisan/petty	12,392	0.168	0.374	11,174	0.164	0.371
Household occupation: salaried	12,392	0.200	0.400	11,174	0.210	0.407
Household occupation: other	12,392	0.050	0.218	11,174	0.050	0.217
Household income quintile 1	12,392	0.210	0.407	11,174	0.178	0.383
Household income quintile 2	12,392	0.21	0.410	11,174	0.196	0.397
Household income quintile 3	12,392	0.211	0.408	11,174	0.214	0.410
Household income quintile 4	12,392	0.195	0.396	11,174	0.230	0.421
Household income quintile 5	12,392	0.169	0.375	11,174	0.182	0.386
Household asset quintile 1	12,392	0.210	0.408	11,174	0.193	0.395
Household asset quintile 2	12,392	0.238	0.426	11,174	0.235	0.424
Household asset quintile 3	12,392	0.219	0.414	11,174	0.211	0.408
Household asset quintile 4	12,392	0.168	0.374	11,174	0.176	0.381
Household asset quintile 5	12,392	0.164	0.370	11,174	0.185	0.388
Household size	12,392	6.615	2.904	11,174	6.180	2.821
Household highest adult education	12,375	6.782	4.896	11,163	7.028	5.014
Membership and media exposure of women in the household (baseline)
Membership of household to women’s group: Mahila Mandal	12,380	0.070	0.255	11,153	0.072	0.262
Membership of household to women’s group: self-help group	12,378	0.098	0.297	11,154	0.093	0.290
Media exposure of women in the household	12,174	0.456	0.332	10,954	0.467	0.333
Conservatism: practice of purdah and men eating first (baseline)
Practice of purdah	10,957	0.578	0.494	9,529	0.570	0.495
Men eating first	10,972	0.332	0.471	9,528	0.328	0.469
Village-level/PSU level characteristics (baseline)
Urban location dummy	12,392	0.331	0.471	11,174	0.347	0.476
Current enrolment of girls (6–16): proportion	12,326	0.819	0.215	10,929	0.854	0.200

Notes: The table includes the sample of young women who were 12–16 years old in 2005, and young men who were 15–19 years old in 2005. The dependent variables are measured using 2012 data from the follow-up survey. All other characteristics are measured using baseline data from the 2005 survey. Mahila Mandals are voluntary service organizations that work for the betterment of women in the villages of India. A self-help group is a community-based group with 10–20 members, usually women, with anti-poverty agendas. The members are from similar social and economic backgrounds, all voluntarily coming together to save small sums of money, on a regular basis.

^α Crime variables are an aggregated measures estimated at village or primary sampling unit level excluding the reporting household. For example, the crime rate for the ith household in jth village is estimated by taking the average of crime rates reported by all other households in the jth village except the ith household.

Source: Author’s compilation based on IHDS data.

Table A3. Regression result: perceived crime in the locality and marriage decision of women with district fixed effects

	Married between surveys		Early marriage
	(1)	(2)	(3)	(4)
	Gender specific crime	All crime	Gender specific crime	All crime
Gender specific crime: harassment of unmarried girls (baseline)	0.075*	0.062	0.011	−0.000
	(0.044)	(0.046)	(0.031)	(0.032)
Theft in the locality (baseline)		−0.012		0.045
		(0.074)		(0.063)
Breaking-in in any household in the locality (baseline)		0.252		0.073
		(0.204)		(0.153)
Threat or attack in the locality (baseline)		0.053		0.034
		(0.114)		(0.085)
Observations	9,799	9,789	9,392	9,382
R-squared	0.318	0.318	0.160	0.160

Notes: Standard errors clustered at the Primary Survey Unit or PSU (i.e. village or town) level are in parantheses.

*** p < 0.01,

** p < 0.05,

* p < 0.1. The sample includes young women who were 12–16 years old in the first survey in 2005 and unmarried at that time. The outcome variables are observed in the second survey round in 2012.

Source: Author’s compilation based on IHDS data.

Table A4. Attrition equation

	Sample attrition
	(1)	(2)
	Gender specific crime	All crime
Gender specific crime: harassment of unmarried girls (baseline)	−0.012	−0.001
	(0.028)	(0.030)
Theft in the locality (baseline)		−0.074
		(0.056)
Breaking-in in any household in the locality (baseline)		0.192
		(0.140)
Threat or attack in the locality (baseline)		−0.057
		(0.080)
Constant	0.378	0.381
	(0.389)	(0.390)
Observations	11,589	11,579
R-squared	0.112	0.112

Notes: Standard errors clustered at the Primary Survey Unit or PSU (i.e. village or town) level are in parantheses.

*** p < 0.01,

** p < 0.05,

* p < 0.1. The sample includes young women who were 12–16 years old in the first survey in 2005 and unmarried at that time. The outcome variable is the attrition from the sample in the follow-up survey. All models include state fixed effects. Control variables include individual characteristics, and household-level and village/PSU-level factors from the baseline survey.

Source: Author’s compilation based on IHDS data.

Table A5. Robustness result: controlling for conservative attitudes of households

	Married between surveys	Early marriage
	(1)	(2)
Gender-specific crime: harassment of unmarried girls (baseline)	0.102***	0.055**
	(0.040)	(0.025)
Purdah dummy	0.030**	0.029***
	(0.014)	(0.010)
Men eat first dummy	0.035***	0.028***
	(0.013)	(0.010)
Observations	8,789	8,437
R-squared	0.258	0.114

Notes: Standard errors clustered at the Primary Survey Unit or PSU level are in parantheses.

*** p < 0.01,

** p < 0.05,

* p < 0.1. The sample includes young women who were 12–16 years old in the first survey in 2005 and unmarried at that time. The outcome variables are observed in the second survey round in 2012. All models include state fixed effects. Control variablesinclude gender-neutral crimes, individual characteristics, and household-level and village-level factors from the baseline.

Source: Author’s compilation based on IHDS data.

Table A6. Robustness result: controlling for village infrastructure and shocks in rural sample

	Married between surveys	Early marriage
	(1)	(2)
Gender-specific crime: harassment of unmarried girls (baseline)	0.140**	0.116***
	(0.059)	(0.043)
Constant	–1.092*	–1.089**
	(0.652)	(0.536)
Observations	6,114	5,840
R-squared	0.244	0.102

Notes: Standard errors clustered at the Primary Survey Unit or PSU level are in parentheses.

*** p < 0.01,

** p < 0.05,

* p < 0.1. The sample is the young women in rural area who were 12–16 years old in the first survey in 2005 and unmarried at that time. The outcome variables are observed in the second survey round in 2012. All models include state fixed effects. Control variables include gender-neutral crimes, individual characteristics, and household-level and village-level factors from baseline survey. The village infrastructural control includes the distances to the nearest police station, bus stop, railway station, and college from the village in kilometres and their square terms. The village shocks include the occurrence of drought, flood, epidemic, earthquake, cyclone, hailstorm, and any other shock in the village between the survey years of 2005 and 2012.

Source: Author’s compilation based on IHDS data.

Table A7. Assessing the robustness of the estimated coefficient accounting for potential bias due to omitted variables

Outcome variable	Uncontrolled coefficient	Controlled coefficient	Identified coefficient (β) for δ = 1 with R^2max = 1.3*R^2controlled	Estimated δ for β = 0 with R^2max = 1.3*R^2controlled
	(1)	(2)	(3)	(4)
Marriage	0.097	0.109	0.112	−18.05
Early marriage	0.054	0.060	0.061	−16.66

Notes: Column (1) presents the estimated coefficient of the main explanatory variable (i.e. gender specific crime) from a regression that does not include any other control variable. Column (2) shows the coefficient from the regression that includes all the control variables and corresponds to the regression presented in column 2 of Table 1. The estimated bias-corrected (identified) coefficient in column (3) and the estimated value of δ in column (4) are obtained using the psacalc command in Stata (Oster, 2019).

Appendix B

Figure B1. Reason for migration.

Notes: The figure includes females 12–16 years old in 2005 and males 15–19 years old in 2005 who migrated between the years 2005 and 2012.

Source: Author’s compilation based on IHDS data.