Policy Research Working Paper 11147 Wealth, Marriage, and Sex Selection Girija Borker Jan Eeckhout Nancy Luke Shantidani Minz Kaivan Munshi Soumya Swaminathan Development Economics A verified reproducibility package for this paper is Development Impact Group available at http://reproducibility.worldbank.org, June 2025 click here for direct access. Policy Research Working Paper 11147 Abstract Two mechanisms have been proposed to explain sex selec- aversion. Marital matching, sex selection, and dowries are tion in India: son preference in which parents desire a male jointly determined in the model, whose implications are heir and daughter aversion in which dowry payments make tested on a representative sample of rural households. Simu- parents worse off with girls. Our model incorporates both lations of the model indicate that existing policies targeting mechanisms, providing micro-foundations, based on the daughter aversion might exacerbate the problem, while organization of the marriage institution, for daughter identifying other policies that could be effective. This paper is a product of the Development Impact Group, Development Economics. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may be contacted at gborker@worldbank.org. A verified reproducibility package for this paper is available at http://reproducibility. worldbank.org, click here for direct access. RESEA CY LI R CH PO TRANSPARENT ANALYSIS S W R R E O KI P NG PA The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Wealth, Marriage, and Sex Selection∗ Girija Borker† Jan Eeckhout‡ Nancy Luke§ Shantidani Minz¶ Kaivan Munshi‖ Soumya Swaminathan∗∗ Keywords. Family Economics. Social Norms. Marriage Market. Sex Selection. Caste. Assortative Matching. Wealth Distribution. Inequality. Control Function. JEL. J12. J16. D31. I3. ∗ Akhil Lohia, Johannes Maywald and Angela Marcela Rojas Franco provided exceptional research assistance and Swapnil Singh graciously helped us with the cross validation. Research support from the National Institutes of Health through grant R01- HD046940, the Keynes Fund at the University of Cambridge, Cambridge-INET, the Agence Nationale de la Recherche (ANR) under the EUR Project ANR-17-EURE-0010, ERC Grant 339186, and grant ECO2015-67655-P is gratefully acknowledged. We are responsible for any errors that may remain. † Development Impact, World Bank & IZA ‡ UPF Barcelona (ICREA-BSE-CREi) § Pennsylvania State University ¶ Christian Medical College, Vellore ‖ Yale University and Toulouse School of Economics ∗∗ M S Swaminathan Research Foundation 1 Introduction Sex selection through female feticide, infanticide, or neglect is a serious problem in many parts of the world, constituting one of the most egregious forms of gender discrimination. Amartya Sen brought sex selection to public attention over 30 years ago when he famously claimed that 100 million women were “missing” in Asia (Sen, 1990). Since that time, India has made tremendous economic progress. Economic development has been associated with greater gender equality on many dimensions (e.g. Geddes and Lueck (2002), Doepke and Tertilt (2009)). However, on the sex selection dimension, the problem has persisted. Child (aged 0-6) sex ratios in India were 109 boys per 100 girls in the most recent 2011 population census, which implies, based on the natural benchmark of 102 set in our analysis, that 2.4 million girls were missing in that age cohort.1 Our analysis sheds light on the underlying motivations for sex selection. Two mechanisms have been proposed to explain this phenomenon in India: son preference in which parents desire a male heir, either to provide old-age support or to inherit their wealth, and daughter aversion, in which marriage payments or dowries make parents worse off with a girl than with a boy (Guilmoto, 2009). The son preference mechanism has been examined theoretically (Edlund, 1999; Bhaskar, 2011) and has also received empirical support (Arokiasamy and Goli, 2012; Bhalotra et al., 2019). In contrast, the daughter aversion mechanism is less well understood, barring some recent evidence that exogenous increases in dowries worsen the sex ratio (Alfano, 2017; Bhalotra et al., 2020). In particular, there is no obvious reason why girls’ families, who are on the buyer’s side of the market when marriage payments flow to the boy, should be necessarily disadvantaged in equilibrium. The first contribution of the model that we develop in this paper is to provide micro-foundations for daughter aversion, or what we call the marriage market mechanism, based on the organization of the marriage institution in India. In our model, we derive specific conditions under which dowries can leave parents worse off with a girl than with a boy leading them to sex select. Parents are endowed with a particular level of wealth that is available for consumption. Parents are altruistic and so if their children were single, they would distribute their wealth so that they and their children consume the same amount. However, this is not the allocation of resources that emerges in equilibrium. Marriage in India is patrilocal, which means that the girl leaves her natal home when she marries. Her parents must thus use the dowry to share wealth with their daughter. The dowry is given to the in-laws and is combined with their own wealth before being subsequently distributed within the marital household. This indirect bequest mechanism gives rise to sex selection under the following conditions: (i) There must be an absence of commitment on the boy’s side. If the girl’s parents could transfer the bequest to her directly or if the boy’s side could commit to making the optimal transfer ex post, then sex selection would not arise. (ii) The girl’s bargaining position in her marital home must be less than that of the boy’s side, which results in her receiving a less than equal share of the available resources. If those resources were divided equally across all members of the household, then there would be no sex selection even in the 1 The natural sex ratio at birth favors boys and is typically assumed to be 105 (Guilmoto, 2009). Subsequent mortality favors girls, but there is no similar widely agreed upon benchmark for the 0-6 age group, which will be the focus of our analysis. As discussed below, widespread sex selection only commenced in the 1980s in South India, which is the setting for our research. Drawing on population censuses prior to that time, we set the benchmark natural sex ratio for children aged 0-6 at 102 boys per 100 girls. 1 absence of commitment. (iii) The social norm in India that all girls must marry must be binding. If that were not the case, a girl’s parents could avoid the disutility associated with the marriage market mechanism by leaving her single.2 The second contribution of the model is to derive the relationship between sex selection and wealth. We do this in two steps: First, we establish that there is positive assortative matching on wealth, which is a well documented feature of the marriage market in India; e.g. Bloch and Rao (2002). The marriage market is precisely defined by the endogamous caste or jati, with many castes coexisting in a given area, and this feature of Indian society will play an important role in the empirical analysis that follows. Given that parents are altruistic, all girls’ parents want their daughters to match with wealthy boys in their caste, whose families have greater resources, thus allowing the daughter-in-law to consume at a higher level. The dowry must, therefore, be increasing steeply enough in wealth to ensure that less wealthy girls’ parents are not willing to deviate from the assortative equilibrium and pay the price that is needed to match up in wealth. The dowry has been characterized in the literature as a price in the marriage market (Becker, 1973) or as a bequest from the girl’s parent to her (Botticini, 1999). A distinguishing feature of the marriage transfer in our model is that it serves both purposes, allowing girls’ parents to make (indirect) bequests, while simultaneously clearing the marriage market (see also Anderson and Bidner (2015)). Next, we establish that sex selection is declining, and the wealth-gap or hypergamy is increasing, as we move down the wealth distribution within a marriage market (caste). The intuition for this result, which is obtained regardless of the mechanism underlying sex selection, is provided by the following example. Suppose that there are 100 wealth levels and two boys and one girl at each wealth level. Except for the number of boys and girls, the wealth distribution is the same and uniform on [1,100]. Then under positive assortative matching, one of the boys with wealth 100 marries the girl with wealth 100 and the other boy marries the girl with wealth 99, the boys with wealth 99 marry the girls with wealth 98 and 97, and so on, until the last boy to be matched, with wealth 50, marries the girl with wealth 1. Hypergamy is increasing as we move down the wealth distribution: at the top (100,100) there is no wealth gap and at the bottom (50,1) the wealth gap is 49. This is obviously not an equilibrium. Poorer parents are less disadvantaged by having a girl since they are marrying up and, thus, they will have less incentive to select the gender of their child. As a result, the sex ratio will adjust and be less biased as we move down the wealth distribution (although the wealth-gap will continue to increase in equilibrium). Once there is sex selection, the wealth distribution on the two sides of the marriage market (which deter- mines the pattern of matching) becomes endogenous. As in the family economics literature; e.g. Greenwood, Guner, and Vandenbroucke (2017), we must account for the two-way interaction between the family’s sex selection decision and the marriage market equilibrium; this implies that sex selection, the wealth distri- bution, and the dowry must be solved simultaneously. The expression for the dowry, in addition, holds a fixed point. This is a challenging problem, which has not been previously solved in the matching literature. We are nevertheless able to show analytically that sex selection exists and is increasing with wealth, under 2 Although social norms have historically received less attention in the literature on marriage in developing countries, there have recently been some attempts to understand how such norms can affect the functioning of the marriage institution (Ashraf et al., 2020; Corno et al., 2020). 2 reasonable parameter restrictions, at the top of the wealth distribution (where the wealthiest boys and girls match with each other) and at the bottom of the wealth distribution (where the dowry of the last boy to match is pinned down by his outside option of remaining single). In addition, we solve the model numerically and observe that sex selection exists and is increasing with wealth at each point in the wealth distribution. Note that the decline in sex selection moving down the wealth distribution, that we have derived, is not driven by an associated decline in equilibrium dowries. While we assume that dowries are always positive, in line with the evidence that dowries are universal in India (Chiplunkar and Weaver, 2023), the model does not generate implications for the relationship between dowries and relative wealth. The reason why sex selection is declining as we move down the wealth distribution is because girls are marrying higher up and because they are in increasingly short supply (which allows the girl’s side to appropriate an increasing share of the marital surplus). The empirical tests of the model use data from the South India Community Health Study (SICHS), which we directed. This study covers a rural population of 1.1 million individuals residing in Vellore district in the South Indian state of Tamil Nadu. The study area is representative of rural Tamil Nadu and rural South India with respect to socioeconomic and demographic characteristics. The analysis makes use of two components of the SICHS: a census of all 298,000 households drawn from 57 castes residing in the study area, completed in 2014, and a detailed survey of 5,000 representative households, completed in 2016. The survey collected information on the marriage of the primary respondent (the household head) and the marriages of his children (in the preceding five years) which we use for the analyses of dowry and hypergamy. To test the predictions of the model for sex selection, we use the SICHS census data, which include nearly 80,000 children aged 0-6. In general, extremely large data sets are needed to detect sex selection with the required level of statistical confidence. The SICHS census is the only data set we are aware of that is large enough to estimate the relationship between household wealth and sex selection within a caste - the social unit that defines a marriage market in India. We begin the empirical analysis by documenting that dowries are always positive, as assumed in the model. Although the model does not have unambiguous implications for the association between dowries and relative wealth, it does predict that dowries given by girls will exceed dowries received by boys at each wealth level (because girls marry up when there is sex selection). We verify this prediction with SICHS survey data and then proceed to test the key implications of the model for hypergamy and sex selection. We test whether hypergamy is increasing (using SICHS survey data) and sex selection is decreasing (using SICHS census data) as we move down the per capita wealth distribution within castes. This result holds when we use average cross-village variation in wealth within caste, but does not hold when using household-level predictors of wealth such as current income, parental education, and landholdings. It is generally believed that wealthy (landowning) households are more likely to practice sex selection (Murthi et al., 1995). There is also anecdotal evidence that this practice is more prevalent in landowning upper castes; e.g. the Jats and Rajputs in North India (Jeffery et al., 1984) and the Gounders and Kallars in South India (George et al., 1992). Recent research that exploits exogenous changes in property rights (Bhalotra et al., 2019) provides causal evidence supporting the postulated association between sex selection 3 and land ownership, which in our model would increase intrinsic son preference. However, this channel is distinct from the implication derived above from our model, and from the models of Edlund (1999) and Bhaskar (2011), which is that sex selection is increasing in relative wealth within the marriage market (defined by the caste in India). Since we have data from multiple castes with different wealth distributions, it is possible to separately identify the relative wealth and absolute wealth effects. Our (conditional) estimates indicate that relative wealth has a positive effect on sex selection, but that absolute wealth has a negative effect on this outcome. This last finding, which runs counter to the conventional view, highlights the importance of distinguishing between relative and absolute wealth. The current child sex ratio in the study area, obtained from the SICHS census, is 109, which is just slightly higher than the corresponding statistic for South India from the 2011 census (108). Although the sex ratio is biased, based on our natural benchmark of 102, it is not exceptional. Nevertheless, within castes, our estimates indicate that the sex ratio varies from 101 in the bottom decile of the wealth distribution, to as high as 118 in the top decile (which is comparable to the sex ratios in the worst states in the country). The marriage market is organized in the same way in all castes and the positive association between sex selection and relative wealth is obtained caste by caste across the social spectrum in our study area. These results highlight the fact that sex selection is not a problem that is restricted to particular areas of the country, as currently assumed, but may in fact be a more pervasive phenomenon that needs to be addressed within each caste. We complete the analysis by estimating the structural parameters of the model and conducting counter- factual simulations. The first simulation decomposes the contribution of the son preference mechanism and the marriage market mechanism: based on our estimates, the latter mechanism accounts for 52% of the variation in sex ratios within castes. The next set of simulations examines the effectiveness of alternative programs that attempt to reduce sex selection by targeting the marriage market mechanism. This analysis is especially relevant, given our new findings on the extent of sex selection within castes. In recent years, the central government and many state governments have introduced cash transfer programs rewarding parents if they have a girl, with the objective of reducing daughter aversion. We find that some existing programs, which target specific (low income) households, could actually worsen sex selection overall, by changing the equilibrium marriage price within castes. While this is an important finding in itself, our counter-factual analysis indicates, in addition, that transfers that flow directly to married women and thus avoid the marriage market channel could be substantially more effective than transfers to parents (even if they are altruistic towards their children). Based on this finding, we argue below that an easily implementable change in existing programs, with regard to the timing of the transfers, could substantially increase their impact. 4 2 A Model of Wealth, Marriage, and Sex Selection 2.1 Marriage in India We take the following features of the marriage institution in India as given in our analysis. First, marriages are endogamous, matching individuals almost exclusively within their caste or jati.3 Second, marriages are patrilocal, with women moving into their husbands’ homes, which are often outside their natal village (Dyson and Moore, 1983; Rahman and Rao, 2004). Third, marriages are arranged by the parents and relatives of the groom and bride, with family wealth being a major consideration when forming a match (Prasad, 1994; Bloch and Rao, 2002; Desai and Andrist, 2010).4 Fourth, the social norm is that all girls must marry (Caldwell et al., 1983; Arnold et al., 1998; Bhat and Halli, 1999; Basu, 1999). Bloch and Rao (2002) note that “getting one’s daughter married is considered an Indian parent’s primary duty and to have an older unmarried daughter is a tremendous misfortune with large social and economic costs.” The consequence of this social stigma is that marriage is universal for women in India, whereas the marriage rate for men has varied by region and over time, depending on the availability of brides, as documented by Gupta (2014), using census data over the course of the twentieth century. Although the level of the dowry is determined endogenously in our model, the way in which the payment is organized is determined by the marriage institution. As discussed, marriages are arranged by the bride and groom’s parents. In addition, more than 90% of new couples begin their married life co-residing with the groom’s parents (Desai and Andrist, 2010). Over time, cash and gold, which can be easily appropriated by the in-laws, have come to constitute an increasing share of the dowry payment (see Prasad (1994) and the references cited therein). Given these features of the marriage institution, it follows that the “dowry is not transferable to the bride, nor does the daughter gain control of the dowry in the way in which the son gains control over land following the partition of his parents’ estate. In fact, even the groom’s control over the dowry is likely to be subordinate to that of his parents as long as the latter are alive” (Sharma, 1980). The fact that the marriage payment flows to the in-laws before it is distributed within the household, will have important consequences for sex selection in our analysis.5 2.2 Model Set Up The model that we develop in this section isolates those elements of the Indian marriage institution that are responsible for sex selection. The model is thus set up to be as parsimonious as possible, abstracting away from many features of the family and the marriage market that are not directly relevant for the analysis. Population. Consider a population of families with measure 2. We assume that a family consists of one 3 Evidence from nationally representative surveys such as the 1999 Rural Economic Development Survey (REDS) and the 2005 India Human Development Survey (IHDS) indicates that over 95% of Indians marry within their caste. Recent genetic analyses have established that these patterns of endogamous marriage have been in place for over 2,000 years (Moorjani et al., 2013). 4 Rao (1993a) finds that household characteristics, especially land wealth, matter more for matching than individual charac- teristics in rural India. Our own results, reported below, indicate that matching on household wealth is independent of matching on education. 5 Bau et al. (2023) consider an alternative allocation of resources in which the son transfers the dowry to his parents, in lieu of old-age support, allowing him (and his spouse) to migrate. Our analysis ignores migration, which remains low in rural India (Munshi and Rosenzweig, 2016). 5 parent and one child. The gender of the parent is irrelevant. The gender of the child is the purpose of our analysis. Under natural circumstances, without sex selection, a child is born a boy or a girl with equal probability, and the distribution of children would each have measure one.6 Families are indexed by their wealth z which is distributed according to the measure Γ(z ) on [z, z ], with Γ(z ) = 2. Denote the boy’s family wealth by x and the girl’s by y . The measure of families with boys and with girls will be endogenous, as will be the distribution of wealth. We denote the wealth distribution of families with boys by F (x) and with girls by G(y ). Under natural circumstances, without sex selection, and with equal probability of having a boy or a girl, the wealth distribution of boys is identical to that of girls: F (·) = G(·) = 1 2 Γ(·). Preferences. Denote the wealth-contingent consumption of parents by Cx , Cy and that of the children by cx , cy . All individuals have logarithmic preferences over consumption, and we assume that families maximize the sum of their members’ utilities U = log(Ci ) + log(ci ) + Ix ub , ∀i = {x, y }.7 Although the focus of the analysis is the marriage market channel for sex selection, we incorporate a coexisting intrinsic son preference channel by assuming that families who have a boy get a utility boost ub . Denote the maximized utility of the groom’s family with wealth x marrying a bride with wealth y by u(x, y ) and the associated utility for the bride’s family by v (x, y ). The Marriage Institution. The model incorporates the key features of the marriage institution listed above. Castes form independent marriage markets and we can think of the model as describing one such market. Marriages are arranged, with family wealth being the major consideration when forming a match. The additional institutional feature that is especially relevant for the model is that marriage in India is patrilocal; i.e. women move into their husbands’ homes. Patrilocal marriage has benefits and costs for the girl’s family. The cost of patrilocal marriage to the girl’s family is that the boy’s parent is only willing to accept the match if the girl’s parent pays a dowry d. The benefit is that the girl will get to consume a fraction of the boy’s family’s wealth (inclusive of the dowry payment). Matching with a well-off boy is thus especially beneficial. Based on the previous description of the marriage payment arrangement, we assume that the boy’s parent retains control of the dowry. The boy’s parent’s total endowment, x + d, is subsequently allocated in three parts: the boy and his (altruistic) parent each receive a share β/2, and the remainder, 1 − β , goes to the girl. The girl’s share reflects her bargaining position in the marital home. We make the standard assumption that her bargaining position is relatively weak; i.e. β ≥ 2/3. We further assume that there is no heterogeneity in the bargaining position, which is determined by the outside option of being single.8 6 Although the sex ratio at birth is not exactly equal to one in practice, what matters for the analysis is the sex ratio in the marriage market. Under natural circumstances, the adult sex ratio is indeed equal to one. 7 Equivalently, parents have altruistic preferences over the utility of their children. The assumption that preferences are logarithmic is broadly consistent with Euler equation estimates of the inter-temporal elasticity of substitution; e.g. Attanasio and Weber (1993), Blundell et al. (1994). Although there are only two generations in the model, we can capture the steady state of a fully dynamic model with overlapping generations by interpreting the weight on the child’s consumption utility as reflecting the cumulative (discounted) weight on all future generations. It is possible that the parent’s and the child’s consumption would then no longer receive equal weight, and this would also be true if parents were not perfectly altruistic, but this extension to the model would not change the results that follow. 8 While higher education will generally improve the girl’s outside options (Anderson and Bidner, 2015), unusually low female labor force participation rates in India weaken this association. Based on the indirect evidence reported below in Section 3, greater education does not improve the girl’s bargaining position in her marital home. In addition, previous studies have 6 Consumption. Given the setup described above, the consumption of all agents (parents and children) of a married groom-bride pair (x, y ) can be written as: β cx = (x + d) 2 β Cx = (x + d) (1) 2 cy = (1 − β )(x + d) Cy = y − d. The dowry d, and the consumption allocations are determined endogenously in the model as functions of wealth, x and y . Matching. Matching in this marriage market is frictionless, with the transfer between the bride and the groom’s family d determined competitively.9 We denote the equilibrium matching allocation by µ(y ), i.e., the family wealth of the groom who is married to a bride with family wealth y is x = µ(y ). The timing of the decisions is as follows: the participants in the marriage market first choose their best partner, given a “Walrasian” schedule of dowries, and the marriage market subsequently clears with a resulting equilibrium price d. For any match between a girl’s family with wealth y and a boy’s family with wealth x, and given a dowry d, the utility of the boy’s family can be written as: β u = ub + 2 log (x + d) . (2) 2 The utility of the girl’s family satisfies: v = log(y − d) + log ((1 − β )(x + d)) . (3) The outside option of the child staying single for the boy’s family is Sx = ub + 2 log x 2 − mb and for the girls’ family: Sy = 2 log y 2 − mg , where mb , mg denote the social cost to the family when a boy, girl stays single. Note that if the child stays single, then the altruistic parent will divide the family wealth equally between the two of them. Note also that son preference applies to the family whether the boy is married or single. Given the social norm that all girls must marry, we assume that the cost of her staying single mg is very high and hence, that the outside option for the girls’ side, Sy → −∞. This difference in options outside marriage for girls and boys will play an important role in generating sex selection below. proposed that the girl’s bargaining position will vary with economic development (Anderson and Bidner, 2015), changes in sex selection technology (Hussam, 2021), and changes in dowry laws (Calvi et al., 2021). These dynamic considerations are outside the scope of our model. 9 We are effectively ignoring delays in matching, which are likely to be small, given that all marriages occur within the tightly integrated caste. We also ignore dowry-based marital violence, which has been associated with asymmetric information in the bargaining between the groom’s family and the bride’s family over the distribution of resources (Bloch and Rao, 2002). 7 2.3 Analytical Solution and Results We solve the model in three steps. First, we show that families on the two sides of the marriage market match on wealth. Second, we show that there is sex selection, which implies a shortage of girls, at the top and the bottom of the wealth distribution. Third, we show that sex selection is increasing in wealth at the top and the bottom of the wealth distribution. Complementing these results, numerical simulations of the model in Section 2.4 show that there is sex selection at each point in the wealth distribution and that sex selection is increasing in wealth across the distribution. Equilibrium Matching. The dowry d is jointly determined with the equilibrium matching allocation, x = µ(y ). In competitive equilibrium, the allocation must be optimal for each agent and the market must clear. In the marriage market, we derive conditions for optimality on the girl’s side, taking as given the maximized utility on the boy’s side, u(x), for each wealth level. Notice that u is now a function of the boy’s family wealth x alone because once the marriage price d has been determined in equilibrium, there will be a distinct price for each wealth level. Using equation (2), we can write the boy’s family wealth as: u−ub 2e 2 x+d= (4) β and as a result, the utility of the girl’s family can be expressed as: u−ub u−ub 2e 2 2e 2 v (x, y, u) = log x + y − + log (1 − β ) . (5) β β A girl’s family with wealth y will take the hedonic Walrasian price schedule, u(x), as given when choosing the partner with wealth x that maximizes its utility given in equation (5). This is a matching problem with Imperfectly Transferable Utility (ITU), as analyzed in Legros and Newman (2007). The first order condition to this problem satisfies vx + vu u′ = 0. (6) Having established the condition for optimality, the remaining condition to be satisfied for a competitive equilibrium is market clearing. We construct the equilibrium allocation x = µ(y ) to ensure market clearing, and then determine the properties of the pattern of matching µ. Denote Mb ≡ exp(mb /2). Proposition 1 There is Positive Assortative Matching on wealth, i.e., µ′ (y ) > 0, provided Mb β < 1. Proof. In Appendix. The condition Mb β < 1 ensures that dowries are positive at every wealth level, as shown below. Given this condition, we establish that there is positive assortative matching and the market will clear from the top, with the wealthiest available girl matching with the wealthiest available boy. Without sex selection, a child is born a boy or girl with equal probability. This implies that the wealth distribution on either side of the market is the same. It follows that girls and boys of equal wealth will match with each other; y = x. If 8 there is sex selection at every wealth level, as derived below, then y = x at the top of the wealth distribution and y < x at all other wealth levels when the market clears from the top. There is no technological complementarity between the boy’s and the girl’s wealth in our model. The complementarity that gives rise to positive sorting is derived from the structure of the marriage institution in conjunction with the parents’ preferences to leave a bequest. Wealthy parents are willing to pay a higher dowry to secure a wealthy match, which will ensure higher consumption for their daughters.10 The first order condition, equation (6), ensures that the hedonic price, u, and, hence, the dowry is increasing sufficiently steeply in x so that the matching on wealth is stable. If the dowry was chosen on the basis of the bequest motive alone, then for a given x, a girl’s family with wealth y would choose d to maximize v (x, y, d) = log(y − d) + log((1 − β )(x + d)). It is straightforward to verify that all girls’ parents would then want them to match with the wealthiest boys and the market would not clear. The dowry thus serves both as a bequest and as a price to clear the marriage market in our model.11 Our assumption that families match on wealth alone is consistent with the empirical evidence that house- hold characteristics, especially land wealth, matter more for matching than individual characteristics in rural India (Rao, 1993a). Our own results, reported below, indicate that matching on household wealth is independent of matching on an important individual characteristic (education). We could add individual characteristics to the model, but this would not generate additional empirical implications and the matching problem then becomes a multi-dimensional allocation problem, which is analytically intractable once the wealth distribution is allowed to be endogenous (on account of sex selection).12 A related consideration motivates the assumption that each family consists of a single parent and a single child in our model. If the family consisted of two parents and multiple children, instead, then sex selection decisions would need to be modeled inter-dependently within the family and we would be faced with a multi-dimensional matching problem once again.13 The advantage of our simplifying assumptions is that they allow us to focus on a single key source of variation – family wealth – in the model, although we will condition on other dimensions of heterogeneity, including education and family size, in the empirical analysis. Sex Selection. Intrinsic son preference and the marriage market mechanism work interchangeably to leave parents better off with boys than with girls; u(y ) − v (y ) > 0 at each wealth level y in equilibrium, as made 10 The implicit assumption in our model is that girls’ parents can transfer any fraction of their wealth as a dowry; i.e. there are no liquidity constraints. The evidence on this assumption is mixed. Recent evidence from India (Anukriti et al., 2022; Corno et al., 2020) indicates that income shocks affect the timing of marriage and that girls’ households change their labor supply leading up to the marriage. However, there is an extensive literature, summarized in Munshi and Rosenzweig (2016), that documents (close to) full risk-sharing in rural India in the face of income shocks that include major contingencies such as illness and marriage. 11 This dual role for the dowry distinguishes our model from existing models of marriage with dowries. In Botticini and Siow (2003) the marriage market clears by wealth matching between brides and grooms, and dowries serve only as a bequest. In Anderson and Bidner (2015), the dowry serves both roles, but two separate instruments are available. Following common convention, we refer to the marriage transfer as the “dowry” throughout the paper. The technically more accurate terminology is that the price component of the transfer is the groom-price and the bequest component is the dowry (Anderson, 2007). 12 Multi-dimensional matching problems are difficult to solve even with exogenous distributions and linear preferences. See, for example, Choo and Siow (2006) and Lindenlaub (2017). 13 Abstracting away from family composition is not uncommon in the economics literature. In research on the labor market, for example, the vast majority of papers assume labor supply and job search decisions are made by individuals independently. However, we know that those decisions are jointly determined within the family. For instance, a spouse may choose to stay at home if his wife has a well-paid job, but not if she is unemployed. 9 precise below. For this gain from having boys to translate into a biased sex ratio, a sex selection technology must be available. We assume that parents who are expecting a girl can replace her with a boy (with probability one) at a utility cost k , which is distributed according to the cumulative density function H (k ).14 k incorporates the monetary cost, which is relatively small, and the more important legal and ethical cost of sex selection. We assume that k is uncorrelated with wealth and is bounded below at zero.15 Although sex selection decisions are made many years before the child enters the marriage market, note that there is no uncertainty in our model. Parents correctly anticipate how the marriage market will clear in the future and, hence, sex selection and matching (with resulting equilibrium dowries and consumption allocations) can be modeled simultaneously. Before proceeding to establish that sex selection is present, we first derive the following result: Lemma 1 Dowries are positive at every wealth level provided Mb β < 1. Proof. To see why this is the case, consider the utility of a married boy’s family when the dowry is zero: u(x) = ub + 2 log βx x 2 . This can be compared with the family’s utility when he is single: ub + 2 log 2 − mb . The family will be better off with the boy remaining single if 2 log β < −mb or, equivalently, Mb β < 1. It follows that if this condition is satisfied, the dowry cannot be zero and, instead, must be strictly positive at every wealth level x for the marriage market to be active. Given Lemma 1, which is consistent with the observation that dowries are universal in India, it can be shown that there will be sex selection at the extremes of the wealth distribution. The matching function, µ(y ), and the dowry, d, are determined simultaneously in equilibrium, together with sex selection. Although this simultaneity prevents us from analytically deriving the level (or presence) of sex selection at each wealth level, this is possible at the extremes because (i) under positive sorting, matching is exogenously determined at the top of the distribution, with the wealthiest girls marrying the wealthiest boys, and (ii) the dowry is pinned down at the bottom of the distribution, where the family of the last boy to match is indifferent between him staying single or marrying. Proposition 2 In equilibrium, there is sex selection at the top and the bottom of the wealth distribution: 1. at the top y = y , for some ub (β ), which is declining in β , whenever ub > ub ; 2. at the bottom y = y , for some mb , whenever mb < mb . Proof. In Appendix. Given that the cost of sex selection, k , is bounded below at zero, we need to establish that u(y ) − v (y ) > 0, in equilibrium, for sex selection to be present at any wealth level y . This will be the case at the top of the 14 In reality, the decision is more subtle. First, if parents use sex selective abortion rather than infanticide or neglect to eliminate unwanted girls, then all parents who anticipate that they will make this decision must bear the ex ante cost of sex determination. Second, even if parents do eliminate a girl, there is no guarantee that the next pregnancy will result in a boy. There is thus a stochastic element to the cost of sex selection that we abstract from in our modeling choice. 15 Given that parents are altruistic, it may be more reasonable to assume that the lower bound for k is strictly positive. This would have no qualitative bearing on the results that follow. 10 wealth distribution if intrinsic son preference, ub , and the boy’s side’s bargaining position, β are sufficiently large. It will be the case at the bottom of the distribution if the cost of staying unmarried for boys, mb , is sufficiently low. In general, to generate sex selection through the marriage market mechanism, we require the following: (i) the absence of commitment (ii) β > 2/3 (iii) mb < mg (which we assume goes to infinity). If the boy’s parent could commit to transferring the dowry in its entirety to the daughter-in-law or β = 2/3, then the girl’s parent would make a marriage payment of y/2 and the girl would end up consuming exactly that amount. The dowry would be increasing in wealth, but there would be no sex selection and, hence, families with exactly the same wealth would match with each other. It follows that we need (i) and (ii) to generate sex selection. We also need (iii), because without the norm that all girls must marry, a girl’s parent could avoid the disutility of having a girl (due to (i) and (ii)) by leaving her single. Once there is sex selection, boys with wealth less than x∗ are left unmatched. If there was bride-price in equilibrium; i.e. payments from boys to girls at the time of marriage, and the population was growing, then the deficit of girls could be cleared by boys “buying” younger girls (Tertilt, 2005; Neelakantan and Tertilt, 2008; Bhaskar, 2011).16 In our model, with dowry payments in equilibrium, the marginal boy that stays single prefers that state to marrying the least wealthy girl (and receiving the dowry that comes with her, conditional on his wealth). If the population is stationary, as it has been in South India since the mid-1990’s, and assuming that the wealth distribution is unchanged in the short-run, that marginal boy will similarly prefer being single to marrying the least wealthy girl in any successive cohort. Thus, the deficit of girls in our model cannot be cleared by allowing boys to match across cohorts.17 Wealth and sex selection. Proposition 2 establishes that there will be sex selection at the top and at the bottom of the wealth distribution if ub , β are sufficiently large and mb is sufficiently small. Numerical simulations reported below show that if these conditions are satisfied, then there will be sex selection at each point in the wealth distribution. We next proceed to describe how the level of sex selection changes with wealth. The example that we constructed above with 100 wealth classes indicates that sex selection will decline in equilibrium as we move down the wealth distribution. The analysis that follows formalizes this argument. With positive assortative matching, girls with family wealth y match with boys with family wealth µ(y ), where dµ(y )/dy > 0. When a family with wealth y that is expecting a girl decides to have a boy instead, it will receive utility u(y ) − k . That boy will then match with a poorer girl with family wealth µ−1 (y ). If the family had chosen instead to keep the girl, it would have received v (µ(y ), y ; u(µ(y ))), which we know from Proposition 2 (and the numerical simulations that follow) is less than u(y ). Thus, the family will proceed with sex selection if its cost k < u(y ) − v (µ(y ), y ; u(µ(y ))). In general, for families with wealth y there is a 16 If the population is stationary, then the age-gap between husbands and wives will widen over successive cohorts until the girls are too young to marry. To illustrate this argument, consider the following thought experiment. Suppose that we are out of steady state and the number of boys is double the number of girls and all boys marry at the age of 25. Then the first cohort of boys will marry girls aged 25 and 24, the second cohort will marry girls aged 24 and 23, and so on. Eventually the girls will be too young and some boys must remain unmarried. This is independent of the sex ratio as long as it is greater than one. 17 Consistent with this argument, we see in Appendix Figure A1, using data from the SICHS census, that the age-gap between spouses has actually narrowed over time (across successive age cohorts). 11 critical cutoff k ⋆ satisfying k ⋆ (y ) = u(y ) − v (µ(y ), y ; u(µ(y ))), (7) where families with k < k ⋆ choose sex selection. Given that the cost of sex selection, k , is distributed according to the cumulative density function, H (k ), the fraction of families with wealth y that choose sex selection is thus H (k ⋆ (y )). For what follows and without loss of generality, we assume H uniform on [0, a]. The pattern of sex selection at every wealth level generates an endogenous and distinct distribution of wealth for girls and boys. The economy-wide distribution of wealth z is Γ(z ). The measure of families with boys whose wealth exceeds x and the measure of families with girls whose wealth exceeds y can thus be described as follows: x y F (x) = (1 + H (k ⋆ (z ))dΓ(z )/2 and G(y ) = (1 − H (k ⋆ (z ))dΓ(z )/2, (8) x y where x = y = z . With Positive Assortative Matching, the market clearing condition is x y dF (x) = dG(y ) (9) µ(y ) y or equivalently: x y (1 + H (k ⋆ (z ))dΓ(z )/2 = (1 − H (k ⋆ (z ))dΓ(z )/2. (10) µ(y ) y Sex selection determines the distribution of wealth for boys and girls, which, in turn, determines the pattern of matching in equation (10). The pattern of matching determines sex selection in equation (7). Sex selection and assortative matching must thus be solved simultaneously. This two-way interaction between a particular family decision and the sorting equilibrium is a common feature of marriage models in the family economics literature. If we knew the payoff u(x) at every level of wealth x on the boys’ side, then we could solve for sex selection and matching recursively, starting at the top of the wealth distribution and moving down. We would know µ(y ) at any wealth level y on the girls’ side, given the pattern of sex selection at higher wealth levels, and so would be able to compute u(y ) − v (µ(y ), y ; u(µ(y ))) and, hence, H (k ⋆ (y )). However, the hedonic price schedule u(x) must also be derived endogenously in the model. To do this we integrate the first order condition in equation (6), vx + vu u′ = 0, which implies u′ = − v vx u , with respect to x: x vx (x, µ−1 (x); u(x)) u(x) = − dx + u(x⋆ ) (11) x⋆ vu (x, µ−1 (x); u(x)) where the denominator is negative, and where x⋆ is the lowest wealth boy who is matched (and is indifferent ⋆ between marrying and staying single). From the outside option, we know that u(x⋆ ) = ub + 2 log x2 − mb . The equilibrium is fully defined by the sex selection condition, the matching condition, and the payoff 12 condition, as specified in equations (7), (10), and (11). This system of equations must be solved simulta- neously. The additional consideration is that the payoff condition holds a fixed point because u(x) appears on both sides of equation (11). We cannot solve the system of equations analytically to establish that sex selection is increasing in wealth at each level. However, the model can be solved numerically (see Section 2.4). We can, moreover, obtain analytical results at the very top of the wealth distribution where the matching pattern is exogenously determined; y = x, and at the lowest wealth level at which boys match, x⋆ , where u(x⋆ ) is determined by the outside option, as specified above. dk⋆ (y ) Proposition 3 Sex selection is increasing in wealth (i.e., dy > 0): 1. at the top y = y , whenever d < y 2; y 2. at the bottom y = y , whenever d(x∗ ) < 2 . Proof. In Appendix. If the result in Proposition 3 holds over the entire wealth distribution, then this implies that sex selection increases monotonically with relative wealth. The intuition for this conjecture, which is validated by the numerical results that follow, is that the shortage of girls grows as we move down the wealth distribution (because more boys are left unmatched above them) which allows their families to retain an increasing share of the marital surplus. In the extreme, at the bottom of the wealth distribution, the family of the last boy to match is pushed down to its outside option; i.e. it is indifferent between him marrying and staying single. The least wealthy girls benefit the most from the surplus of boys and, hence, sex selection is lowest at the bottom of the wealth distribution.18 2.4 Numerical Solution and Results The algorithm. The numerical solution of the model assumes that there is a finite number of wealth classes. This implies that boys and girls in a given wealth class could potentially match across multiple wealth classes. The matching allocation then looks like a step function instead of a smooth curve. With a dv continuum of wealth classes, the first order condition in equation (6), dx = 0, ensures that the allocation and transfers are optimal for girls’ families in each wealth class. With a finite number of wealth classes, the equivalent condition is that girls’ families in a given wealth class will obtain the same utility across all the wealth classes that they match with. Given that the equilibrium payoff for the boys’ families, u(x), is a function of their wealth alone, the symmetric condition is that boys’ families in a given wealth class receive the same utility across all the wealth classes that they match with. The solution to the model must satisfy the sex selection condition, the measure preserving allocation or matching condition, and the payoff condition simultaneously. The algorithm that we use to solve the model 18 In our model, relatively less wealthy girls are advantaged because they marry up and are on the short side of the marriage market (which improves their family’s ex ante bargaining position). An alternative explanation for the relative wealth effect is based on these girls having stronger ex post bargaining positions in their marital household because relatively less wealthy girls marry at a later age or have a smaller age-gap with their spouse. As observed in Appendix Figures A2 and A3, there is no empirical support for these alternative explanations. 13 numerically begins with an initial guess for the payoff at the top of the wealth distribution, u(x), and for the pattern of matching. We know from Proposition 2 that there will be sex selection at the top and the bottom of the wealth distribution. This implies that there will be a shortage of girls in the highest wealth class and so girls in the next to highest wealth class will match up (with boys one wealth level higher than themselves) and horizontally (with boys in their own wealth class). As we move down the wealth distribution, the excess of boys accumulates and it is possible that below some wealth level, girls match exclusively with wealthier boys. Given any initial guess for u(x) and the matching pattern, we can solve for u(x) and v (y ) in each wealth class. v (y ) is a function of y , x, and u(x), as specified in equation (5). Given that girls in the highest and the next to highest wealth class match with the wealthiest boys, with family wealth x and payoff u(x), we can solve for v in both wealth classes. Girls’ families in the next to highest wealth class must receive the same utility, v , from matching with the wealthiest boys and boys in their own wealth class. This allows us to solve for u in the next to highest wealth class. We continue to solve recursively in this way down the wealth distribution. With sex selection, boys below a wealth level x⋆ will remain unmatched. A comparison of u(x⋆ ) derived in x⋆ the first iteration with the outside option, ub + 2 log 2 − mb , is used to adjust the guess for u(x) in the next iteration. Given u and v derived in the first iteration, the level of sex selection H (k ∗ (y )) can be determined in each wealth class y . The pattern of matching implied by this sex selection is used as the starting point for the next iteration. This iterative process continues until there is convergence. The numerical solution thus simultaneously satisfies the sex selection condition, the matching condition, and the payoff condition.19 Numerical simulations. The wealth distribution is assumed to be log-normal in the numerical simulations, with the parameters selected to match the SICHS census data (within castes). The wealth distribution is divided into 100 classes for the simulations. As assumed above, k ∼ U [0, a], which implies that there are four parameters in the model: β , ub , mb and a. We select values for these parameters: β = 0.8, ub = 0.7, mb = 0.12, and a = 23 that are in line with the values estimated below. The matching pattern generated by the model is reported in Figure 1a. Notice that the plot is not a smooth function and has small steps. This is due to the discreteness of the wealth distribution, which results in each wealth class matching with multiple wealth classes of the opposite sex as described above. At higher wealth classes, girls and boys match horizontally as well as up and down, respectively. This is why the plot touches the 45 degree line at those wealth levels. However, below a certain wealth level, girls match exclusively with wealthier boys, shifting the plot above and away from the 45 degree line. Hypergamy, or the wealth-gap, increases as we move down the wealth distribution because of the growing stock of unmatched boys, with the intercept of the plot measuring the fraction of boys that end up being single. We can compute the dowry that boys of a given wealth class x receive directly from the value of u(x) β derived for that wealth class. Recall that u(x) = ub + 2 log 2 (x + d) . While this is the same, regardless of the wealth of the girls’ families that they match with, the dowry paid by girls in a given wealth class 19 Although we cannot establish analytically that the equilibrium is unique, we verified that the algorithm generated the same allocations in practice with different starting conditions. 14 will vary with the wealth of the families that they match with. It is thus necessary to take account of the matching pattern in each wealth class when computing the average dowry paid by girls’ families over the wealth distribution. Given that girls are matching up on average, the model unambiguously predicts that the dowry given must be greater than the dowry received at each wealth level. This is indeed what we observe in Figure 1b. Moreover, the dowry is positive at all wealth levels if the condition specified in Lemma 1 is satisfied and this is also true in our numerical simulation. Note that the model does not have a clear prediction for how the dowry will vary across the wealth distribution. As we move down the distribution, girls match with increasingly wealthy boys, but their family’s share of the marital surplus is also increasing. The first effect shifts up the dowry, whereas the second effect works in the opposite direction and, hence, the net effect on the dowry is ambiguous. We see in Figure 1b that the dowry (as a ratio of household wealth) is initially decreasing and then increasing in wealth. Figure 1: Simulated Model (a) Marital Matching (b) Sex Selection and Dowry 90 100 .49 .5 .485 80 .4 40 50 60 70 Dowry/ Wealth Male wealth class .48 Pr(Girl) .3 .475 30 .2 .47 20 10 .465 .1 0 0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100 Wealth class Female wealth class Received Given We have shown analytically that there will be sex selection at the top and the bottom of the wealth distribution (Proposition 2) and that sex selection is increasing in wealth at the extremes (Proposition 3). In contrast with the model’s ambiguous implications for the dowry, we expect sex selection to be monotonically declining as we move down the wealth distribution because girls are in increasingly short supply (and their families thus receive an increasing share of the marital surplus). This is indeed what we observe in Figure 1b. Notice that the dowry is less than half the family’s wealth at the top and the bottom of the wealth distribution. This satisfies the condition for sex selection to be increasing in wealth at both points, as derived in Proposition 3. 15 3 Empirical Analysis 3.1 Descriptive Evidence Demographic and socioeconomic characteristics. The South India Community Health Study (SICHS) covers a rural population of 1.1 million individuals residing in Vellore district in the state of Tamil Nadu. There are 298,000 households drawn from 57 castes in the study area. The study area is representative of rural Tamil Nadu (with a population of 37 million) and rural South India (comprising Tamil Nadu, Andhra Pradesh, Karnataka, and Maharashtra with a total population of 193 million) with respect to demographic and socioeconomic characteristics.20 Appendix Table A1, Panel A, reports demographic and socioeconomic characteristics, separately for males and females, in the study area, rural Tamil Nadu, and rural South India. Statistics for Tamil Nadu and South India are based on official Government of India data, while the corresponding statistics for the study area are derived from the SICHS census. The age distribution, marriage rates, literacy rates, and labor force participation rates are comparable between the study area and both rural Tamil Nadu and rural South India. Appendix Table A1, Panel B compares the religious composition across the three populations. Over 90% of these populations is Hindu and thus our characterization of the marriage institution in the model, based on Hindu social norms, applies to almost the entire population. We complete the descriptive analysis in Panel B by examining overall and child (aged 0-6) sex ratios. Overall sex ratios in the population are close to parity, which can be explained by the fact that sex selection in South India is a relatively recent phenomenon and because life expectancy is greater for females than males.21 In contrast, child sex ratios, which are comparable across the three populations, and range from 107 to 109, are clearly above the natural benchmark, which we define as 102 boys per hundred girls, based on statistics for South India from the 1961 and 1971 rounds of the population census. The benchmark we have chosen reflects the history of sex selection in South India. Marriages in this region were traditionally between close-kin (Dyson and Moore, 1983). The most preferred match for a girl was her mother’s younger brother or, if he was unavailable, one of her mother’s brothers’ sons (Kapadia, 1995). There were no major payments at the time of marriage, just a ritual gift or stridhan from the groom’s side to the girl (Srinivas, 1989; Anderson, 2007). The commitment problem on the boy’s side, which gives rise to sex selection in our model, was eliminated in two ways: (i) Since the two families (dynasties) sequentially traded girls across the generations, the long-term (repeated) nature of the relationship allowed cooperation to be sustained. (ii) Given the extremely close pre-existing relationship between the girl’s natal family and her husband’s family, the two families effectively functioned as a cooperative unit. Parents were not disadvantaged by having a girl in this marriage system and, as a result, there was no sex selection. Caldwell et al. (1983) and Srinivas (1984) attribute the demise of this system to economic development and the resulting changes in wealth within 20 The SICHS was designed to examine a variety of socioeconomic phenomena and health problems, including the treatment of tuberculosis. The study area thus comprises three Tuberculosis Units (TU’s) within Vellore district that were purposefully selected to be representative of rural South India. Dyson and Moore (1983) define the South Indian region by the same set of states. Kerala is excluded from the list of South Indian states because it is an outlier on many socioeconomic characteristics. 21 Sex selection only commenced in South India in the early 1980’s, as described below, and so the first cohorts to be affected would be in their mid-30’s when SICHS data collection was completed in 2016. 16 castes. Families that had traded girls over many generations no longer had the same level of wealth. Close- kin marriage declined (Caldwell et al., 1983; Kapadia, 1993) and a marriage market consequently emerged to match unrelated families within the caste on wealth, with a marriage price or dowry clearing the market.22 By the 1980’s, the practice of dowry was observed across the caste distribution in South India (Caldwell et al., 1983). The widespread emergence of dowries in South India in the early 1980’s coincided with the onset of sex selection, which is why sex ratios prior to that point in time can be assumed to be unbiased.23 Marriage patterns. The analysis in this paper makes use of two components of the SICHS: a census of all households and a detailed survey of 5,000 households who are representative of the castes in the study area.24 The survey collected information on key aspects of the marriage institution: (i) whether marriage was within the caste, (ii) whether marriage was between close-kin, (iii) whether the marriage was arranged, and (iv) whether the female spouse was born in a different village. This information was collected from the primary respondent (household head) for his own marriage and for the marriages of his children in the five years preceding the survey. Table 1: Marriage Patterns Generation Parents Children Balanced sample Males Females P-value Males Females P-value (1) (2) (3) (4) (5) (6) (7) Same Caste 0.98 0.95 0.93 0.22 0.92 0.89 0.33 Related 0.48 0.33 0.35 0.51 0.30 0.26 0.36 Arranged 0.86 0.76 0.89 0.74 0.74 0.81 0.24 Female moved outside natal village 0.75 0.71 0.81 0.65 0.81 0.81 0.32 Mean dowry [As a fraction of annual income] 172.51 199.72 0.00 132.44 208.43 0.01 [3.16] [4.15] [1.20] [2.87] Minimum dowry (in thousand Rupees) 0.0 0.0 0.0 0.0 Maximum dowry (in thousand Rupees) 1,117.2 1,417.2 533.9 655.2 Observations 3,524 421 611 53 53 Source: SICHS household survey Table 1 provides information on marriages over the two generations based on data from the SICHS survey; for the parents in Column 1 and for their children (who married in the past five years) in Columns 2-3. Notice that the number of sons who married is lower than the corresponding number of daughters. This is because women marry younger than men in India, as documented with SICHS census data in Appendix Figure A4. Although most women marry in their twenties, men will marry into their thirties. Some men who marry in their thirties will have fathers in their sixties. Given that the fathers (household heads) in the survey are aged 25-60, older boys are under-represented in Column 2. We account for this in Table 1 and Table 2 that follows, 22 In related research, Anderson (2003) links dowry inflation to economic development and increased income inequality on the male side of the marriage market. 23 It is possible that sex selection due to son preference predates the 1980’s in which case we will underestimate the extent of sex selection. 24 The sampling frame for the household survey included all ever-married men aged 25-60 in the SICHS census plus (a small number of) divorced or widowed women with “missing” husbands who would have been aged 25-60, based on the average age-gap between husbands and wives. The sample was subsequently drawn to be representative of each caste in the study area, excluding castes with less than 100 households in the census. 17 by reweighting the sample of male marriages such that the age distribution matches the age distribution of the girls’ husbands (which is not subject to sample selection). As an alternative, more stringent basis for comparison between males and females, we restrict the sample to households in which both a son and a daughter married in Columns 5-6. While the discussion that follows is based on the full set of marriages, we note that our findings hold up with the balanced (restricted) sample of brothers and sisters.25 Consistent with nationally representative survey evidence and genetic evidence for the country as a whole, we see in Table 1 that 98% of the parents and 95% of the children married within their caste. The incidence of close-kin marriage declines, in line with the general trend in South India described above, from 48% in the parents’ generation to 35% in the current generation. However, most marriages continue to be arranged. Girls moved from their natal village in a substantial fraction of the marriages. We will take advantage of this feature of the marriage institution in India, by exploiting information on the natal villages, to test the model’s predictions for marital matching below. Table 1 also reports the dowry in levels and as a fraction of the household’s annual income, for the marriages of the children that took place in the last five years. The dowry amount is computed by summing up the monetary value of gifts, such as household items, vehicles, and gold, as well as the cost of the wedding celebration. The annual income is measured by the profit in the past year from land owned, leased, or rented plus the wage earnings of all adult members. Dowries in South India are now as high as they are in the North (Caldwell et al., 1983; Srinivas, 1984; Rahman and Rao, 2004; Anderson, 2007) and in line with past studies; e.g. Rao (1993b), Jejeebhoy and Sathar (2001), Rahman and Rao (2004), and Chiplunkar and Weaver (2023) the dowry is 3 – 4 times the household’s annual income on average, which is a substantial sum in an economy where access to market credit is severely restricted.26 Notice that dowries paid by girls are larger than dowries received by boys. This observation is consistent with our model when there is sex selection because girls will then marry wealthier boys on average. Hypergamy results in girls paying a higher dowry than boys receive at each level of wealth, as documented in our theoretical simulations in Figure 1b. The gender difference in dowries is also observed with the restricted sample of marriages in Columns 5-6; sisters pay a (significantly) higher dowry than their brothers receive. An alternative explanation for the dowry-gap is social desirability bias, with girls’ parents reporting more than they gave and boys’ parents under-reporting the amount that they received. We cannot rule out such bias, but note that it will have no bearing on the analysis of hypergamy, based on marital matching, that follows. We complete the description of marriage patterns in Table 1 by documenting that the minimum payment given and received is equal to zero, consistent with our assumption in the model that dowries are always positive. Marital Matching. Table 2 provides descriptive evidence indicative of both assortative matching and hypergamy, based on the SICHS survey data. The survey respondents were asked whether their child’s spouse’s family had the same wealth, more wealth, or less wealth than their own. The majority of marriages, 25 Five households in this restricted sample have two siblings of the same gender. For these households, we drew a sibling at random from the over-represented gender when constructing the statistics that are reported in Table 1 and Table 2. 26 Most households will receive support from their close relatives and other caste members to pay the dowry. Munshi and Rosenzweig (2016) use data from the Rural Economic and Development Survey (REDS) to document that gifts and loans within the caste are the primary source of support for meeting major contingencies, including marriage, in rural India. 18 Table 2: Marital Matching Sex of the child Full sample Balanced sample Males Females Brothers Sisters (1) (2) (3) (4) Partner’s parental household Wealthier 0.09 0.18 0.08 0.13 Same wealth 0.65 0.64 0.69 0.68 Less wealthy 0.26 0.17 0.24 0.19 Kolmogorov-Smirnov test of equality p-value = 0.001 p-value = 0.998 Observations 421 611 53 53 Notes: The sample in Columns 1-2 consists of children who were married in the past five years. The sample in Columns 3-4 is restricted to households in which at least one girl and one boy married in the past five years. For the five households in the restricted sample with two siblings of the same gender, we chose a random sibling from the over-represented gender. The statistics in Column 1 are re-weighted such that the age distribution of the sons matches the age distribution of the daughters’ husbands. The Kolmogorov-Smirnov p-value is from a test of the equality of the distribution of responses. Source: SICHS household survey. for males and females, are reported to be with families of equal wealth, consistent with assortative matching. At the same time, the respondents are more likely to report that their daughters married up in wealth than their sons, which is indicative of hypergamy. Conversely, they are more likely to report that their sons married down in wealth than their daughters. The Kolmogorov-Smirnov test easily rejects the null hypothesis that the distribution of responses is equal for males and females. Qualitatively similar results are obtained with the balanced (restricted) sample, which compares siblings within households, although the sample size is too small to reject the hypothesis that the distribution of responses for brothers and sisters is statistically equal. While the statistics in Table 2 are based on coarse categories, we will provide evidence of both assortative matching and hypergamy with a finer measure of wealth in Section 3.2. Upper caste marriages in North India have long been associated with hypergamy (Bhat and Halli, 1999). Hypergamy has also been associated with the emergence of dowry in South India (Caldwell et al., 1983; Srinivas, 1984). These are all settings with sex selection. However, previous studies have failed to make the connection between hypergamy and sex selection. Indeed, given that marriages are almost exclusively within the caste, girls cannot marry up on average without sex selection. Sex selection. Although the model assumes that each parent has a single child, a couple will usually have multiple children. Going beyond the model, we posit that the marriage market channel will bias the sex ratio of children at all birth orders. In contrast, if parents have access to a relatively certain sex selection technology and they want a single male heir, either to support them in old age or to inherit their wealth (property), then they will postpone sex selection until they are close to their desired family size. In particular, first births will not be biased due to the son preference channel as long as parents want at least two children, which is typically the case in rural India. Appendix Table A2 reports child (aged 0-6) sex ratios from three sources: the 2005-2006 and 2015-2016 19 rounds of the Demographic and Health Survey (DHS) and our own SICHS census, which was conducted in 2012-2014, between the two DHS rounds. All three data sources include the birth order of each child and, hence, it is possible to compute sex ratios for first-born children and for all children. We see that the sex ratios are very similar across the three data sets. The sex ratio for first born children (105–107) is elevated relative to the natural sex ratio of 102, indicating that the marriage market channel is active.27 Moreover, the sex ratio for later-born children (110-112) is larger than for first-born children, implying that the son preference channel is also present. If we assume that the marriage market mechanism affects the sex ratio at all birth orders equally, while the son preference mechanism only operates at higher birth orders, then 55% of sex selection in the study area can be attributed to the marriage market mechanism. 3.2 Testing the Model Dowries and Relative Wealth. To examine the association between dowries and relative wealth we need to construct a measure of the latter variable. The first challenge when constructing this measure is that household wealth is not directly observed. The SICHS census and survey both collected information on the household’s income in the preceding year.28 We purge these incomes of their transitory component, which will be substantial in a rural (agrarian) economy, by taking the average of the incomes reported in the census and the survey. This gives us a measure of permanent income that can be used as a proxy for household wealth. Ideally, we would like to average over multiple income realizations to remove the transitory component completely, but this limitation will not undermine the key facts that we establish below. The second challenge when constructing a measure of relative wealth is that a household does not consist of a single parent and a single child in practice. In our model, the efficient allocation of family wealth is to divide it equally between the parent and the child; i.e. the child (boy or girl) gets half of the family wealth. In a household with multiple children, the equivalent dowry (bequest) for each child will be based on per capita wealth. We account for this in the empirical analysis by constructing measures of per capita wealth; i.e. household wealth divided by household size.29 The latter is a choice variable and we address the possibility that fertility and sex selection are jointly determined below. The final step when constructing our measure of relative wealth is to specify the set of comparison households. Although the marriage market in our model is based on a single cohort, the age-gap between partners will vary in practice. In lieu of a clear partition of age cohorts into independent marriage markets 27 This observation does not necessarily contradict the common assumption that the sex ratio at birth for first-born children is unbiased in India. Jayachandran and Kuziemko (2011) and Jayachandran and Pande (2017) document differential nutritional inputs by gender and birth order among children in India, which could elevate the child (aged 0-6) sex ratio even if the sex ratio at birth is unbiased. 28 Household income is measured by the profit from land owned, leased, or rented plus the total labor income of all members, including those that have temporarily migrated to work. Profit is measured over the entire year, whereas labor income is measured in the month prior to data collection (and then scaled up to the annual level). 29 Households consisting of a single couple and their children, but possibly including other adults (typically a grandparent) account for 96.2% of all households with (co-resident) children in the census. The empirical analysis is based on these households and per capita wealth is computed by dividing household wealth by the number of family members; i.e. the two parents plus their children. The implicit assumption is that other adults in the household do not receive a share of the wealth. For example, a grandfather living with his son’s family would have already distributed his wealth among his children. The results that follow are robust to constructing per capita wealth on the basis of household size. 20 within the caste, we construct a benchmark measure of relative wealth that is based on the household’s position in its caste’s per capita wealth distribution. The implicit assumption with this measure is that the distribution of per capita wealth within the caste is stable across age cohorts. In addition, we construct an alternative measure of relative wealth, which is based on the sample of households that is used for the dowry analysis; i.e. those households who report the marriage of a child in the past five years. Although our benchmark measure is based on a definition of the marriage market that may be too expansive and the alternative measure may be based on a definition that is too narrow, the results below are similar with both measures. Figure 2a reports the relationship between dowries and relative wealth. The estimation sample consists of all marriages of the primary respondents’ children that took place in the five years preceding the SICHS survey. When the child is a girl, the dowry is based on the amount that was given, and when the child is a boy, the dowry is based on the amount that was received. To account for the fact that these values will vary for boys and girls with the same family wealth, as described in Figure 1b, we include a gender dummy in the estimating equation. We also include a full set of caste dummies to account for the fact that some castes will be wealthier than others and thus report higher dowries on average.30 The gender dummy and the caste fixed effects are partialled out using the Robinson (1988) procedure prior to nonparametric estimation. The same procedure is used in all the nonparametric regressions that follow.31 Dowries in Figure 2a are measured by their absolute value, rather than as the ratio with respect to household wealth, as in Figure 1b, because the latter variable also appears on the right hand side of the estimating equation. This implies that our estimates cannot be directly compared with the simulation results in Figure 1b; recall that dowries (as a ratio of household wealth) were initially declining and then increasing in that figure. Estimates with the benchmark measure of relative wealth, based on the entire caste’s wealth distribution, are reported in blue and estimates with the alternative measure, based on a restricted sample of households are reported in red (with accompanying 95% confidence intervals). Dowries are monotonically increasing in relative wealth, with the alternative measures generating very similar estimates. Although the model does not have unambiguous implications for the association between dowries and wealth, it does tell us that dowries given by the girl’s side will exceed dowries given by the boy’s side at each relative wealth level when there is sex selection (and resulting hypergamy). This is indeed the case in Figure 2b, where we report the association between dowries and relative wealth, separately for boys and girls. The amount given by the girls exceeds the amount received by the boys at each wealth level, with little overlap in the confidence intervals, matching the gender-gap in dowries that we reported in Table 1. Marital Matching and Relative Wealth. The model predicts that there will be positive assortative matching on wealth within each marriage market (caste). With sex selection, the additional implication is that 30 Recall from the model that the level of the dowry is pinned down by the outside option of the last boy to match, which is increasing in his family wealth. 31 Consider the estimating equation: yi = f (Zi ) + Xi β + ϵi , where yi is an outcome such as the dowry for household i, Zi is the household’s relative wealth, Xi is a vector of covariates that needs to be partialled out prior to nonparametric estimation of the yi − Zi association and ϵi is a mean-zero disturbance term. The Robinson procedure is implemented as follows: Step 1. Separately regress yi and each element of the Xi vector nonparametrically on Zi . Step 2. Regress the residuals from the first equation, ξˆy , on the residuals from the other equations, ξˆX , using a linear specification without a constant term to estimate β ˆ. Step 3. Nonparametrically regress yi − (Xi − X )β ˆ on Zi , where X is the sample mean of each element in the vector of covariates. 21 Figure 2: Dowry (a) Dowry and Relative Wealth (b) Dowry and Relative Wealth by Gender 300 Dowry (in thousands of Rupees) Dowry (in thousands of Rupees) 300 200 100 200 100 0 .2 .4 .6 .8 1 Rank in caste wealth distribution 0 .2 .4 .6 .8 1 Rank in caste wealth distribution Rank based on households with married children Rank based on all households in the caste Males Females Notes: This figure plots the nonparametric relationship between the dowry and relative wealth, after partialling out the gender dummy and caste fixed effects. In Panel (a), this relationship is based on the household’s position in its caste’s per capita wealth distribution (in blue) and among households in the caste with married children (in red). Panel (b) plots the relationship separately for boys and girls. Source: SICHS household survey girls will marry up, with the wealth-gap or hypergamy increasing as we move down the wealth distribution within a caste. These implications were described graphically in Figure 1a. While we presented descriptive evidence that is consistent with positive assortative matching and hy- pergamy in Section 3.1, this evidence was based on coarse wealth categories. We now proceed to test the model’s implications for matching with finer grained wealth data. The SICHS survey provides information on the household head’s marriage and the marriages of his children that occurred in the five years preceding the survey. Although the spouses’ household wealth is not observed, we do know the identity of their natal villages. We use this information to construct a measure of household wealth, on both sides of the marriage market, as described below. Historical records on agricultural productivity for each village in the SICHS study area and the surround- ing region are stored in the British Library in London. A notable feature of this research project is that we digitized these records and then merged them with the SICHS data. There are 377 panchayats or village governments in the SICHS study area. These panchayats were historically single villages, which subsequently divided or added new habitations. The panchayat as a whole, which often consists of multiple modern villages, can thus be linked back to a single historical village. The data we collected from the British Library includes information on the agricultural revenue tax, per acre of cultivated land, that was levied on each village in 1871 by the British colonial government. This tax assessment was based on the potential output per acre, which, in turn, was determined by exogenous geo-climatic characteristics such as soil quality, crop suitability, temperature, and rainfall. While we expect these fixed characteristics to determine household wealth or permanent income; i.e. the wealth flow today, we also expect that this relationship will vary across castes, which have distinct land ownership and occupation structures within a village. To construct a predictor of 22 household wealth, we thus estimate the following equation: yijk = δj + γk + Rk · δj + ϵijk (12) where yijk is the income for household i belonging to caste j and residing in village k , obtained from the SICHS census, Rk is the 1871 tax revenue in that village, δj is a vector of caste dummies, γk is a vector of village dummies, and ϵijk is a mean-zero disturbance term. Predicted income from this estimated equation provides us with a measure of household wealth at the village-caste level. The F-statistic measuring joint significance of the δj , γk , Rk · δj terms in the prediction equation (12) is 30.05, indicating that fixed conditions at the caste-village level are strong determinants of current household income in this rural economy.32 Based on the estimated relationship described above, we can construct a consistent predictor of wealth for the surveyed household and the spouse’s household (based on its caste and the historical tax revenue in its natal village).33 The wealth measures that we construct for the two households in a match will typically not be the same because, as observed in Table 1, 80% of women move from their natal village when they marry. Table 3, Column 1 reports estimates from an equation in which the male partner’s relative wealth is the dependent variable and the female partner’s relative wealth is the independent variable. The household heads in the SICHS survey range in age from 25 to 60. The younger heads are thus born around the same time as the children of the oldest heads. To increase statistical power, our sample for the analysis of marital matching thus includes both generations, with the restriction that the included household heads be born after 1980; these birth cohorts would have been subjected to sex selection, even in South India, with subsequent gender imbalance in the marriage market. Recall that this gender imbalance is a prerequisite for hypergamy. The coefficient on the female’s relative wealth is positive and significant, which tells us that there is positive assortative matching on wealth. However, it is significantly smaller than one, which is indicative of hypergamy from Figure 1a. The intercept is also positive and statistically significant, as observed in that figure when there is hypergamy. Table 3, Column 2 includes the female’s education, measured by her rank among female members of her caste who were born in a 10-year window around her birth year, as an additional regressor.34 The coefficient on the education variable is small in magnitude and statistically insignificant. A comparison of Columns 1 and 2 indicates that the relative wealth coefficient is hardly affected by the introduction of female education, supporting the assumption in the model that matching on family wealth is independent of individual characteristics. To provide further support for this assumption, we replace the relative wealth of the boy’s household by his relative education as the dependent variable in Table 3, Column 3. The coefficient on his partner’s relative household wealth is now small in magnitude and statistically insignificant, whereas the coefficient on her relative education is positive and significant. Matching on household wealth appears to 32 Standard errors are clustered at the caste-village level when estimating the predicting equation, with the sample restricted to castes with at least 10 households. 33 The 1871 tax revenue is available for all villages in the northern Tamil Nadu region that were directly taxed by the colonial government. The estimated equation can thus be used to predict current wealth even for villages outside the study area. 34 The relative education level is computed with respect to women from the caste in the same age range in the SICHS census. This is because the number of observations in the SICHS survey declines substantially once we go down to the caste-gender-age level. 23 Table 3: Marital Matching Dependent variable Relative wealth of groom Relative education of groom (1) (2) (3) Rel. wealth of bride 0.542 0.542 -0.055 (0.039) (0.039) (0.031) Rel. educ. of bride – 0.009 0.479 (0.032) (0.032) Constant 0.225 0.221 0.420 (0.023) (0.028) (0.024) Observations 699 699 699 Notes: Sample restricted to household heads born after 1980 and children who married in the past five years. Relative wealth measured by rank in the caste wealth distribution, from 0 (poorest) to 1 (wealthiest). Education measured relative to all females/males in the SICHS census in the same caste who are no more than five years younger or older. Standard errors are clustered at the caste-panchayat level. Source: SICHS household survey. be independent of matching on individual characteristics, at least with respect to one important characteristic: education. Given the level of sex selection in the study area, and the resulting excess of boys, the intercept in Table 3, Column 1 should be approximately 0.1. The estimated constant term is too large and, consequently, the estimated relative wealth coefficient is too small. One explanation for this discrepancy is measurement error in the relative wealth of the bride, which will bias the associated coefficient downward (and the intercept upward). This is very likely to be the case since our wealth measure does not vary across households within a village-caste. A second explanation is based on the fact that we have not adjusted for family size, which is not observed for spouses’ households, and thus have not constructed a measure of per capita wealth for the marital matching analysis. While the magnitude of the estimated coefficients should thus be treated with caution, it is nevertheless quite striking that variation in wealth across villages can be used to uncover assortative matching and marital hypergamy within castes. Sex Selection and Relative Wealth. The central prediction of the model is that sex selection will increase as we move up the wealth distribution within castes. Large samples are needed to identify sex selection with the requisite level of confidence. For the analysis of sex selection within castes, we thus turn to the SICHS census data; as observed in Table A2 there are nearly 80,000 children aged 0-6 in the study area. The dependent variable in the equation that we use to test for sex selection is a binary variable indicating whether a child is a girl. The key explanatory variable is the child’s household’s position in the caste per capita wealth distribution. Per capita wealth in Table 4, Column 1 is constructed as the household’s income, obtained from the SICHS census, divided by family size. We also include caste fixed effects in the estimating equation to account for caste-based norms governing the (social) cost of sex selection and village fixed effects to allow for spatial variation in access to sex selection technology and the economic returns to boys versus girls. The coefficient on relative per capita wealth is negative, but very close to zero.35 One explanation 35 We use the Linear Probability model to estimate the relationship between the sex of the child and relative wealth for ease of 24 for this finding is that a single income realization, from the year preceding the SICHS census, measures the household’s wealth with error. To address this potential problem, we replace the wealth measure with a principal component of parental education, per capita income, and landholdings in Column 2. The coefficient on relative per capita wealth is negative but very close to zero. Column 3 uses predicted household income, where the prediction equation includes caste fixed effects, village fixed effects, and 1871 village tax revenue and its interaction with the caste effects, as in equation (12), to measure household wealth. Recall that we used the same measure in Table 3 to study marital matching. The association between the girl dummy and relative per capita wealth, based on this measure, is reported in Table 4, Column 3. The relative wealth coefficient is now negative and statistically significant, as implied by the model. The analysis that follows will subject this core result to further scrutiny by (i) disentangling the relative wealth and absolute wealth effect, and (ii) by verifying that it is robust to alternative construction of the relative wealth measure and different partitions of the sample. We will conclude the analysis in this section with a discussion on omitted variable bias and the control function strategy that we employ to estimate the causal effect of relative wealth on sex selection. Table 4: Sex Selection Dependent variable Girl dummy Measure of household wealth Reported income Principal Component Predicted income (1) (2) (3) (4) Relative wealth -0.00595 -0.00013 -0.05882 -0.06186 (0.00892) (0.00701) (0.01033) (0.01042) Household wealth – – – 0.01533 (0.00672) Mean of dependent variable 0.480 0.480 0.480 0.480 Observations 78,872 78,872 78,872 78,872 Caste FE Yes Yes Yes Yes Village FE Yes Yes Yes Yes Notes: Sample restricted to children aged 0-6 years. Relative wealth measured by rank in the caste per capita wealth distribution, from 0 (poorest) to 1 (wealthiest). Household wealth in Column 1 is measured by reported monthly income, obtained from the SICHS census. Wealth in Columns 2 is a principal component of parental education, per capita monthly income, and landholdings (in acres). Father’s and mother’s education is measured as their highest level of education, in years. Wealth in Columns 3-4 is derived from a predicting equation that includes 1871 tax revenue in the village and its interaction with caste effects along with caste and village (panchayat) fixed effects. Bootstrapped standard errors (in parentheses) are clustered at the caste - panchayat level. Source: SICHS census. We begin in Table 4, Column 4 by adding (predicted) household wealth to the estimating equation. In general, the effect of household wealth on sex selection in a rural population is ambiguous. Landowning households will have a greater demand for a male heir, but wealthy parents are less dependent on their children for old-age support. Household wealth could also be associated with other independent determinants of sex selection; for example, wealthier households could be better educated or have better access to sex selection technology. In our data, we see that the household wealth coefficient is positive and significant, while the interpretation because the mean of the dependent variable is close to 0.5. 25 negative coefficient on relative per capita wealth remains significant and is relatively stable, going from Column 3 to Column 4.36 Figure 3: Sex Selection (a) Benchmark Specification (b) Augmented Specification .6 .6 .55 .55 Pr(Girl) Pr(Girl) .5 .5 .45 .45 0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1 Rank in caste wealth distribution Rank in caste wealth distribution Rank based on all households in the caste Benchmark specification Rank based on households with atleast one child aged 0-6 years With control function Notes: This figure plots the nonparametric relationship between probability that a child is a girl and relative wealth, after partialling out caste and village fixed effects. In Panel (a), this relationship is based on the household’s position in its caste’s per capita wealth distribution (in blue) and among households in the caste with at least one child aged 0-6 (in red). Panel (b) plots the relationship after also partialling out the optimal control function (in red), with the benchmark plot in blue. Source: SICHS census Figure 3a reports nonparametric estimates corresponding to Table 4, Column 3. The household’s position in its caste’s per capita wealth distribution in the benchmark specification, shown in blue with the accom- panying 95% confidence interval, is based on all households in the caste. The alternative construction of relative wealth, shown in red, is based only on households in the caste with at least one child aged 0-6. This is a very different set of households, but the negative relationship between the probability that a child is a girl and relative wealth continues to be obtained. When constructing per capita wealth, household wealth must be divided by family size. We have an accurate measure of family size in the dowry analysis because the children getting married are adults and even if they have younger siblings, the fertility of their mothers will be complete.37 This is not the case for the sex selection analysis, which is based on children aged 0-6. We account for this in Appendix Figure A5a by separately estimating the association between the probability that a child is a girl and relative wealth for children aged 0-6 and 7-13. Family size will be more accurately measured for the older cohort and while we would not want to compare the two cohorts directly, it is reassuring to observe that the same negative 36 One way to circumvent the inefficient dowry-based bequest mechanism would be for girls’ parents to invest in their education, since human capital cannot be as easily misappropriated. If that were the case, then we would expect educated parents, who are more easily able to invest in their children’s education, to take advantage of this alternative strategy (and, hence, have more girls). As reported in Appendix Table A3, the sons and the daughters of more educated parents are more likely to be enrolled in higher secondary school. However, sex selection, measured by the probability that a child aged 0-6 is a girl, is independent of parental education. This finding is also consistent with the assumption (discussed in the model section) that more educated girls do not have greater bargaining positions in their marital homes. If they did, then more educated parents, with more educated daughters, would be more likely to have a girl. 37 Very few families in the study area have more than three children and birth-spacing rarely exceeds five years. 26 association is obtained. Once we extend the model to allow for multiple children in the empirical analysis, we must also account for the accompanying variation in birth order and its consequences for sex selection. As discussed in the previous section, we expect the marriage market mechanism to be relevant at all birth orders and the intrinsic son preference mechanism to be relevant at higher birth orders. Both mechanisms imply that sex selection will be increasing in relative wealth, and this is what we observe in Appendix Figure A5b, for first-born children and later-born children. Moreover, the sex ratio is less biased for first-born children, consistent with the statistics reported in Table A2, because the son preference channel is absent for those children. Sex selection in our model is determined by intrinsic son preference and the marriage market channel. However, other determinants of sex selection could coexist with these mechanisms. For example, the increas- ingly biased sex ratios over time in India have been attributed to a number of factors including: (i) improved sex selection technologies (Arnold et al., 2002; Bhalotra and Cochrane, 2010); (ii) changes in the economic returns to having boys versus girls (Rosenzweig and Schultz, 1982; Foster and Rosenzweig, 2001); and (iii) reduced fertility coupled with a desire to have at least one son (Basu, 1999; Jayachandran, 2017).38 The same factors could generate cross-sectional variation in sex selection. To understand the biases that could be generated by these independent determinants of sex selection, consider the following estimating equation: P (yij = 1) = f (wij , w−ij ) + ξij , where yij = 1 if child i in caste j is a girl and 0 if a boy, wij is the per capita wealth of the child’s family, and w−ij is a vector representing the per capita wealth of all other families in the caste-based marriage market. The f (wij , w−ij ) function determines the family’s relative wealth and the ξij term collects the remaining independent determinants of sex selection. Access to sex selection technology and the differential economic returns to boys versus girls could potentially vary with household wealth, Wij . A reduction in family size, Nij , worsens the sex ratio when there is a demand for a single male heir. In addition, Wij and Nij could be Wij correlated with other independent determinants of sex selection. Per capita wealth, wij = Nij and, hence, E (ξij | wij , w−ij , δj ) ̸= 0, biasing the estimated relationship between the sex of the child and relative per capita wealth. Our solution to this identification problem is to include a flexible control function, g (Wij , Nij ), in the estimating equation. This solution differs from the approach of including a limited number of control variables, which are often imperfect proxies for the omitted variables, in the estimating equation. The key to our strategy is the observation that the determinants of sex selection listed above and, more generally, any independent household-specific determinant, can only bias our estimates of the relative wealth effect if they are correlated with ωij , through Wij or Nij .39 38 As noted, sex ratios will worsen at higher birth-orders when there is a demand for a male heir. Exogenous fertility decline makes the sex ratios worsen earlier, resulting in an overall increase in the bias. This argument has been used to explain why programs that couple incentives to reduce fertility and to have daughters could result in a worsening of the sex ratio (Anukriti, 2018). 39 As in Table 4, we also include caste effects and village effects in the estimating equation to account for factors that determine 27 It is not standard practice to use the control function approach, first proposed by Heckman (1976) in the context of selection correction, to deal with omitted variable bias. This is because the objective with this approach is to subsume the entire component of the error term, ξij , that is correlated with the running variable, ωij , in a single control function, and this is typically infeasible when there are multiple omitted variables. The control function approach is especially effective in dealing with selection bias because there is a single decision that determines this bias. A polynomial function of a single propensity score (estimated in a first-stage selection equation) can thus be used to extract the component of the error term that is associated with the selection bias (Heckman and Navarro-Lozano, 2004). The control function approach can be used to correct for omitted variable bias in our application because: (i) any household-specific determinant of sex selection can only bias our estimates through Wij , Nij , (ii) there are multiple marriage markets organized at the level of the caste, and (iii) the wealth distribution varies across castes. Once the g (Wij , Nij ) function is included in the estimating equation, we are effectively comparing the sex selection decisions of families with the same wealth, Wij , and size, Nij , but who have different relative per capita wealth because they belong to castes with different wealth distributions. When using instrumental variables to deal with omitted variable bias, the objective is to find a variable that is (mean) independent of the error term. This rules out the use of choice variables as instruments because they will typically be correlated with other unobserved determinants of the outcome of interest. With the control function approach, the objective is to capture the component of the error term that is correlated with the running variable in its entirety. Independence is not required with this approach (Heckman and Navarro- Lozano, 2004), and choice variables – family size in our application – and outcomes – household wealth in our application – can be included in the control function without undermining the identification strategy. The challenge in this case is to correctly specify the control function: it should not be too parsimonious (under-fitting the data) nor should it be too flexible (over-fitting the data). Very few families in the study area have more than three children and, hence, household size, Nij , is specified as a vector of binary variables indicating whether household i in caste j has 3, 4, or 5 (and above) members. We allow the control function g (Wij , Nij ) to be linear, quadratic, cubic, or quartic in household wealth, Wij . In addition, we allow for a wide range of interactions between Wij and Nij in the control function. For example, when g (Wij , Nij ) is specified to be a quadratic function of Wij , we allow for (i) no interactions between the family size dummies and the wealth terms, (ii) interactions with the linear wealth term, and (iii) interactions with both the linear and quadratic wealth terms. This leaves us with 14 possible specifications of the control function. Figure A6 reports the relationship between the probability that a child is a girl and relative wealth with each of these specifications, using the Robinson procedure to partial out the terms in the control function and the caste and village fixed effects, as usual. There is little variation in the estimated relationship across the alternative specifications of the control function and the probability that the child is a girl continues to be declining in relative wealth (with three exceptions). Das et al. (2003) recommend cross validation to determine the optimal specification of the control function when correcting nonparametrically for sample selection and endogeneity. We use k -fold cross validation, with sex selection at the caste and village level. 28 k = 10, to partition the sample of children aged 0-6 into 10 randomly selected groups (stratified by caste and relative wealth within castes to maintain balance). Of the 10 subsamples, a single subsample is retained for testing; i.e. to compute the forecast error, while the model is estimated with the remaining subsamples. This process is repeated 10 times, with each testing subsample, to compute the average forecast error with each specification of the control function. These average forecast errors are reported in Appendix Table A4, where we see that the optimal specification of the control function, which minimizes the forecast error, is linear in wealth and without interactions. Figure 3b reports the relationship between the probability that a child is a girl and relative wealth with the optimal control function (partialled out) in red. The corresponding relationship without the control function, which is also reported in Figure 3a, is in blue. Appendix Table A5 reports linear regressions, with the estimated relative wealth coefficients, corresponding to the nonparametric plots in Figure 3. The advantage of including multiple castes in the analysis is that a control function can be included in the estimating equation, which allows us to isolate the association between sex selection and relative wealth within the caste. To independently assess the robustness of our results, we nonparametrically estimate the association between sex selection and relative wealth, caste by caste. The marriage market is organized the same way in all castes, and thus we would expect the predictions of the model to apply to all castes. Figure A7a reports this test for the 0-6 age group for the 12 largest castes in our study area, which account for 82% of the children in this age group. The probability that a child is a girl is decreasing with relative wealth for 9 of the 12 castes. For the three castes that it is not, the number of observations is relatively small (less than 2,000 children per caste). It is possible that the anomalous pattern for these castes is simply a consequence of the small sample size, which makes the estimated relationship unstable. To examine this possibility, we report the association between relative wealth and the probability that the child is a girl for the 7-13 year olds in Figure A7b. Reassuringly, the relationship is negative at each point in the wealth distribution for the three castes for which anomalous patterns were observed in Figure A7a. 4 Structural Estimation and Quantification We complete the analysis in this paper by estimating the structural parameters of the model. The model can subsequently be used to (i) quantify the contribution of the alternative mechanisms to sex selection, and (ii) to evaluate the effect of existing and counterfactual policies targeting sex selection. The relative wealth measure is a continuous variable taking values between 0 and 1. For the structural estimation, we partition this measure into 10 equally sized wealth classes within each caste. The number of classes is chosen by weighting two competing considerations: The larger the number of wealth classes, the closer we can approximate the nonparametric plots that we have used to describe the sex ratio at each point in the relative wealth distribution. However, this comes at the cost of less precise estimates of the sex ratio within wealth classes, which is the unit of observation for the current analysis. To estimate the structural parameters with these data, we utilize the same algorithm that was used to solve the model numerically (for given parameter values) in Section 2. We solve the model, using this algorithm, for all combinations of the parameters to find the combination for which the predicted fraction of girls across the ten equal-sized 29 wealth classes matches most closely with the actual fractions; i.e. for which the sum of squared errors is minimized. We use the 12 largest castes for the structural estimation. (Block) bootstrapped means and confidence intervals for the parameters are reported in Table A6 by drawing repeated samples (of castes) with replacement. The parameter associated with boys’ bargaining position, β , is equal to 0.73, satisfying the condition in the model that its value should exceed 2/3 for the marriage market mechanism to be active. The son preference parameter, ub , is also positive and statistically significant, indicating that this mechanism for sex selection is also present. The parameters associated with the cost to boys from being single, mb , and the cost of sex selection, a, both have the expected (positive) sign and are precisely estimated. The estimated structural parameters can be used to decompose the contributions of the marriage market channel and the son preference channel to sex selection. The lower surface of the blue area in Figure 4a traces out the predicted fraction of girls, based on the estimated model, for each wealth class. The line dividing the red and blue areas traces out the corresponding fraction of girls when the son preference channel is shut down; i.e. ub is set to zero. The red area thus measures the contribution of the marriage market channel, while the blue area measures the contribution of the son preference channel, under the assumption that the fraction of girls equals 0.5 without sex selection. Under this decomposition, 52% of sex selection in the study area is due to the marriage market channel, which is close to the figure of 55% that we arrived at in Section 3.1, based on the sex ratios of first-born and later-born children in Table A2. Figure 4: Quantification (a) Sex Selection Mechanisms (b) Counterfactual Policy Experiments .49 .5 .485 .49 Pr(Girl) Pr(Girl) .48 .48 .475 .47 .47 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 Wealth class Wealth class No transfer Transfer to girls families Marriage market Son Preference Transfer to poor girl families Transfer directly to girls Source: SICHS census Converting the fraction of girls to the number of boys per 100 girls, the official statistic used by the government of India, the sex ratio ranges from 101 to 118 across the ten wealth classes in our data. To put these statistics in perspective, the sex ratio derived from the 2011 population census ranges from 115 to 120 in the three worst states in the country (all of which are North Indian states). This indicates that the variation in sex ratios within castes in the study area is as large as the variation across states in the country. There 30 is nothing unusual about the study area with respect to demographic and socioeconomic characteristics. Sex selection may thus be more serious and more pervasive than is commonly believed, affecting relatively wealthy households within their castes throughout the country. Although our analysis provides new evidence on the extent of sex selection and the mechanisms underlying this phenomenon, the problem itself is well known and widely discussed in academic and policy circles. Many states and the central government have responded to the problem by introducing Conditional Cash Transfer schemes with the stated objective of improving the survival and the welfare of girls and reversing the bias in the child sex ratio. Sekher (2010) evaluates 15 such schemes. These schemes have a number of common features. Parents receive a cash transfer when (i) the birth of a female child is registered, (ii) she receives the requisite immunizations, and (iii) she achieves specific educational milestones. In addition, an insurance cover is provided, which matures when the girl turns 18. Although governmental transfers when she is young go directly to her parents, the insurance payment, when it matures, goes into a bank account that is set up for the girl. Most girls in India are unmarried at 18 and are thus dependents in their parents’ home when the (large) final transfer is received by the family. It thus seems reasonable to assume that the transfer is appropriated by the girl’s parents; this transfer coincides with the time when many girls enter the marriage market and, thus, it can be conveniently used to pay the dowry. Our model is well suited to examine the impact of these policies since it incorporates the son preference and the marriage market mechanism, and both mechanisms have been seen to contribute substantially to sex selection. The dowry is also determined endogenously, allowing for general equilibrium effects. In the framework of our model, a transfer to the girl’s parents is equivalent to an exogenous increase in the wealth of the girl’s family. As an additional counter-factual policy exercise, we examine an alternative scheme that provides a transfer to the daughter after she is married. We have argued above that the entire dowry payment, including the bequest component, is appropriated by the in-laws and then redistributed. One concern is that government transfers to married women would be similarly appropriated. However, recent evidence from rural India indicates that direct payments into married women’s bank accounts can, in fact, be controlled by them (Field et al., 2021). The alternative transfer scheme targeting married women may thus be feasible. Modifying equation (5) in the model, the schemes described above will change the maximized utility of the girl’s family in the following ways: (a) If the wealth transfer, w, is to the girl’s parents, u−ub u−ub 2e 2 2e 2 v = log x + y + w − + log (1 − β ) . (13) β β (b) If the wealth transfer is directly to the girl, u−ub u−ub 2e 2 2e 2 v = log x + y − + log (1 − β ) +w . (14) β β If the maximized utility on the boy’s side, u, is fixed, then the most effective scheme will target the family member; i.e. the girl or her parent, who has a lower level of consumption in equilibrium. However, the effect 31 of the wealth transfer is more complex than that because it will change the equilibrium marriage price and, hence, matching and sex selection over the entire wealth distribution. This is especially important because while some of the welfare schemes are available to all families with girls, many are targeted to families below the poverty line. While the targeted families may be induced to have more girls, there will be spillover effects through the equilibrium marriage market price that could increase sex selection at other points in the wealth distribution. The grey solid line in Figure 4b is the benchmark fraction of girls predicted by the model in each wealth class. The first counter-factual policy experiment that we consider, whose effect is described by the green dashed line, is a 20% wealth transfer to families in the bottom two classes with girls. This experiment is designed to reflect the wealth eligibility requirement in many existing schemes. The fraction of girls increases substantially in each of the two treated wealth classes. This increase in the number of girls at the bottom of the wealth distribution will shift the entire equilibrium price (dowry) schedule and we see in the figure that this results in an increasingly biased sex ratio in the upper eight wealth classes.40 This result would not be obtained with models of sex selection that ignore the pecuniary spillovers within castes, highlighting the value of the market equilibrium analysis for policy evaluation. The next policy experiment that we consider, whose effect is described by the grey dashed line in Figure 4b, provides the wealth transfer to all girls’ parents. To be comparable with the first experiment, the amount of the per family transfer is divided by five (because the beneficiaries are now in 10 rather than 2 of the equal-sized wealth classes). The transfer marginally affects the fraction of girls in each wealth class because part of the subsidy is transferred to boys’ families via higher dowries. The final policy experiment that we consider, whose effect is described by the green solid line, has the most promise. It is the same as the preceding experiment, except that the subsidy goes directly to the adult girls rather than their parents. Note that this transfer should not be given until the girl is married and it cannot be used as a dowry payment. As we can see in Figure 4b, there is now a substantial increase in the fraction of girls in each wealth class.41 This is because the bequest that must be transferred to the girl through the inefficient dowry mechanism will decline. With the resulting decline in the mismatch between the girl’s actual consumption and the preferred level of consumption from her parent’s perspective, it is less costly to have a girl. Policies that give resources directly to girls when they are adults, as opposed to their (altruistic) parents when they are children, may thus be especially effective in reducing the bias in child sex ratios in India. 40 The wealth increase of the girls at the bottom pushes up the dowry for them, but also for those who do not receive the subsidy, since they compete for the same boys. The higher equilibrium dowry leads to more sex selection further up the wealth distribution. 41 This policy experiment is conducted holding constant the β parameter. It is possible that the girl’s bargaining position will increase when she has direct control of the resources she brings into the marriage. The resulting decline in β will further increase the fraction of girls in equilibrium. 32 5 Conclusion Two mechanisms have been proposed to explain the persistently high rates of sex selection in India: son preference in which parents desire a male heir and daughter aversion in which dowry payments make parents worse off with girls than with boys. The model developed in this paper (incorporating both mechanisms) provides micro-foundations, based on the organization of the marriage institution in India, for daughter aversion. We show that sex selection due to daughter aversion arises when the following conditions are satisfied: (i) the in-laws cannot commit to transferring the bequest component of the marriage payment to the girl; (ii) the daughter-in-law receives a less than equal share of the resources available for consumption in her marital home; and (iii) the social norm that all girls marry is binding, or else girls’ parents could avoid the disutility associated with the marriage market channel by leaving their daughters single. While the preceding conditions may hold in the Indian context, they may not elsewhere, providing an explanation for why daughter aversion is not observed in other countries. Take Bangladesh, for example. Marriage in Bangladesh is patrilocal, dowries have emerged in recent years (Ambrus et al., 2010), and there is pressure on girls to marry (Field and Ambrus, 2008). What makes Bangladesh different is an institution – the mehr – that exists in all Islamic societies and which guarantees the woman a share of her husband’s family’s wealth in the event of a divorce. In the context of our model, the mehr improves the outside option for married women, increasing their bargaining position. This might explain why the emergence of dowries in Bangladesh did not increase sex selection (Kabeer et al., 2014), in contrast with what was observed in South India. Next, consider China, where there is a long tradition of son preference and the accompanying deficit of girls goes back many centuries, but where the daughter aversion mechanism appears to be absent. Indeed, increasingly biased sex ratios have been accompanied by a substantial increase in bride prices. Based on our model, we argue that this outcome has emerged in China because (i) the married couple rarely cohabits with the in-laws and, hence, the bequest to the girl and the marriage payment from the boy’s parents to the girl’s parents can be decoupled (Zhang and Chan, 1999); (ii) most Chinese women work and, hence, the bride is in a position to ensure that she receives the bequest.42 Sex selection (generated by daughter aversion and by sone preference), marital matching, and dowries are jointly determined in our model. We show that under reasonable parametric restrictions, sex selection is generated at every wealth level within the marriage market, which is defined by the caste in India. In addition, hypergamy (the wealth-gap between grooms and brides) is increasing, and sex selection is decreasing, as we move down the wealth distribution. These implications are verified with unique data from the South India Community Health Study (SICHS) covering a representative sample of rural households. A novel nonparametric (cross validated) control function approach is used to estimate the relative wealth effect, independently of other determinants of sex selection, effectively comparing outcomes for households with the same (absolute) wealth and family size who happen to be located at different positions in their caste’s (per capita) wealth distribution. Although the overall child sex ratio in the study area is not unusually biased and is comparable to 42 Based on official ILO statistics, female labor force participation in 1990 and 2017, respectively, was 73% and 61% in China and 35% and 27% in India. 33 the corresponding statistic for rural South India, we document wide variation in the sex ratio across the wealth distribution within castes (comparable in magnitude to the variation across states in the country). Our decomposition analysis, based on the estimated model, indicates that about 48% of this variation in sex ratios can be attributed to the daughter aversion (marriage market) mechanism, with the remainder accounted for by the son preference mechanism. Our subsequent counter-factual analyses, based once again on the estimated model, examine the impact of alternative programs targeting sex selection. In recent years, the government has implemented a number of conditional cash transfer programs, rewarding parents if they have a girl, with the stated objective of reducing sex selection. Our analysis indicates that these programs, particularly those restricted to less wealthy households, could actually worsen the problem on net through the general equilibrium effect on marriage prices (dowries). In addition, our analysis indicates that a small, easily implementable, change in current programs – targeting the transfer to married women directly rather than to their (altruistic) parents earlier in life – could substantially increase their impact. 34 Appendix A Tables and Figures Table A1: Comparison of Demographic and Socioeconomic Characteristics Men Women Region South India Tamil Nadu Study Area South India Tamil Nadu Study Area Panel A Age distribution married (%) <10Yrs 0.0 0.0 0.0 0.0 0.0 0.0 10-14 0.0 0.0 0.0 0.0 0.0 0.0 15-19 0.2 1.5 0.0 0.5 0.0 0.8 20-24 2.2 6.9 1.0 5.2 1.0 5.7 25-29 5.4 8.0 4.4 7.9 4.7 7.8 30-34 6.7 7.2 6.8 7.3 6.4 6.4 35-39 6.8 6.7 6.6 6.6 6.9 7.5 40-44 6.2 5.7 6.3 5.9 6.2 5.5 45-49 5.5 5.0 5.9 5.2 6.6 5.6 50-54 4.6 3.4 5.0 3.6 4.7 3.4 55-59 3.6 3.4 4.1 3.8 4.0 3.2 60-64 2.9 2.1 3.4 2.0 3.7 1.9 65-69 2.2 1.3 2.5 1.2 2.3 1.1 70-74 1.4 0.7 1.5 0.5 1.5 0.3 75-79 0.8 0.3 0.9 0.2 0.7 0.1 80-84 0.4 0.1 0.4 0.1 0.4 0.0 85> 0.2 0.1 0.2 0.1 0.2 0.0 Kolmogorov-Smirnov 1.00 1.00 - 1.00 1.00 - test of equality (p-value) Literacy rate (%) 78.5 82.0 86.8 61.2 65.0 69.7 Labor force participation 79.8 81.1 81.0 44.9 42.6 40.0 rate (%, 15-59 years) Panel B Hindu(%) 91.0 92.9 93.7 Muslim(%) 5.3 2.6 1.8 Christian(%) 1.3 4.2 4.5 Other(%) 2.4 0.2 0.0 Sex Ratio 102 101 100 Child Sex Ratio 107 107 108 Population 97,390,696 18,679,065 474,384 95,226,008 18,550,525 475,022 Notes : South India includes rural Maharashtra, Karnataka, Andra Pradesh, and Tamil Nadu. Tamil Nadu refers to rural Tamil Nadu. % married measures the number of married individuals in each age category as a fraction of the total population, seperately for men and women. Literacy rate is defined by the Goverment of India as the percentage of those aged 7+ who can, with understanding, read and write a short, simple statement on their everyday life; SICHS Census definition is those aged 7+ with ≥ 1 year of education (figures for ≥ 3 years of education are similar, 73.8% for men and 59.5% for women). Labor force participation is defined as the proportion of 15-59 year old persons of the total 15-59 years population who are either employed or seeking or available for employment. Sex Ratio refers to the number of males per 100 females in the population. Child sex ratio is the number of males per 100 females for those aged between 0-6 years. Sources: For Tamil Nadu and South India, the data on age distribution, literacy rate, religious composition and sex ratios are from the 2011 Census of India. The data on labor force participation is from the Ministry of Labor and Employment, Government of India, 2009-10. For Study Area, all statistics based on SICHS Census. 35 Table A2: Sex Ratios Population Rural South India Rural Vellore DHS DHS SICHS census Data Source 2005-06 2015-16 2012-14 t-test p-value (1) (2) (3) (4) (5) First-born children 105 107 106 0.344 0.765 Later-born children 111 112 110 0.805 0.880 All children 109 110 108 0.780 0.764 t-test p-value 0.146 0.188 0.045 Observations 5,750 27,072 79,027 Notes : Sex ratios are computed for children aged 0-6 as the number of boys per 100 girls. The t-test in Columns 1-3 reports the p-value of a two sample unpaired t-test for equality of the sex ratio for first-born and later-born children. Column 4 reports the p-value for equality of the sex ratios in DHS 2005-06 and the SICHS data and Column 5 reports the p-value for equality of the sex ratios in DHS 2015-16 and the SICHS data. Table A3: School Enrollment and Sex Selection Dependent variable Boys higher secondary Girls higher secondary Girl dummy enrollment enrollment Age range 14-17 14-17 0-6 (1) (2) (3) Mother’s educations 0.0106*** 0.0126*** -0.000422 (0.000608) (0.000612) (0.000509) Father’s education 0.0112*** 0.0109*** 0.000167 (0.000595) (0.000600) (0.000510) Mean of dependent variable 0.838 0.842 0.479 Observations 27,103 25,300 91,403 Caste FE Yes Yes Yes Notes: Higher secondary enrollment indicates whether the child is enrolled in school. The lower bound for the age range is set at 14 because most children in rural Tamil Nadu study till the 8th grade (age 13). The upper bound is set at 17 because girls start to marry (and leave their parental homes) by the age of 18. Sex selection is measured by the probability that the child (aged 0-6) is a girl. *** p<0.01. Source: SICHS census 36 Table A4: Mean Forecast Sum of Square Errors Linear Quadratic Cubic Quartic (1) (2) (3) (4) No interaction 0.025148 0.025163 0.025154 0.025163 Linear interaction 0.025230 0.025241 0.025239 0.025252 Linear + quadratic interaction 0.025197 0.025163 0.025180 Linear + quadratic + cubic interaction 0.025661 0.025641 All interactions 0.025384 Notes. Mean forecast error is based on k-fold cross validation, with k=10. We consider 14 specifications of the control function: linear, quadratic, cubic and quartic functions of wealth, with varying degrees of interaction between the wealth terms and the family size dummies. All specifications include caste and village fixed effects. Table A5: Linear Regressions Corresponding to Each Specification in Figure 3 Dependent variable Girl dummy Specification Figure 3a (blue) Figure 3a (red) Figure 3b (red) (1) (2) (3) Relative wealth -0.108 -0.0553 -0.0574 (0.00732) (0.00710) (0.00732) Mean of dependent variable 0.499 0.479 0.497 Observations 79027 79068 79027 Caste FE Yes Yes Yes Village FE Yes Yes Yes Absolute wealth No No Yes Family size FE No No Yes Notes: This table reports the linear regression corresponding to each specification in Figure 3. Caste and village fixed effects are first partialled out, using the Robinson procedure, in all columns. The optimal control function is also partialled out in Column 3. Standard errors (in parentheses) are clustered at the caste-panchayat level. Source: SICHS census. Table A6: Structural Parameter Estimates Mean [95% Confidence Interval] Boys’ bargaining position: β 0.730 [0.722,0.739] Son Preference: u0 0.515 [0.446,0.584] Boys’ cost of being single: mb 0.362 [0.324,0.401] Cost of sex selection: a 23.470 [22.921,24.019] Notes: Sample restricted to children aged 0-6 in 12 largest castes. Each caste is partitioned into 10 equal sized wealth classes and the sex ratio is computed within each class. Bootstrapped confidence intervals in brackets. Source: SICHS census 37 Figure A1: Age Difference Between Husband and Wife by Birth Cohorts 12 Age difference between husband and wife (years) 2 4 6 8 10 1940 1960 1980 2000 Husband's birth cohort Notes: This figure shows the age-gap (in years) between spouses with respect to the husband’s birth cohort. Source: SICHS census 38 Figure A2: Variation in Age at Marriage by Relative Wealth 2826 Age at marriage 24 22 20 0 .2 .4 .6 .8 1 Rank in caste wealth distribution Males Females Notes: This figure shows the age at marriage for men and women with respect to relative wealth. Source: SICHS survey Figure A3: Variation in Spousal Age Gap by Relative Wealth 12 10 Age gap between spouses (years) 2 4 6 8 0 .2 .4 .6 .8 1 Rank in caste wealth distribution Notes: This figure shows the age gap (in years) between spouses with respect to relative wealth. Source: SICHS census 39 Figure A4: Proportion Married by Age 1 .8 Proportion married .4 .2 0 .6 20 25 30 35 40 Age Males Females Notes: This figure shows the proportion of men and women who are married at each age (18-40 years). Source: SICHS census Figure A5: Sex Selection across Groups (a) Across age groups (b) Across birth order .7 .7 .6 .6 Pr(Girl) Pr(Girl) .5 .5 .4 .3 .4 0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1 Rank in caste wealth distribution Rank in caste wealth distribution Under 6 First born 7-13 year old Later borns Notes: This figure shows the nonparametric relationship between probability that a child is a girl and relative wealth, after partialling out caste and village fixed effects. In Panel (a), children aged 0-6 years are shown in red and children aged 7-13 years are shown in blue, with the accompanying 95% confidence interval. In Panel (b), the first-born are shown in red and later born children are shown in blue. Source: SICHS census 40 Figure A6: Sex Selection and Relative Wealth - Control Function .55 .5 Pr (girl) .45 .4 0 .2 .4 .6 .8 1 Rank in caste wealth distribution Notes: This figure shows the nonparametric relationship between probability that a child is a girl and relative wealth, after partialling out caste and village fixed effects and different specifications of the control function. Each line depicts one of the 14 specifications of the control function, as explained in Table A4. The three control function specifications that include cubic interactions of predicted income with family size dummies exhibit an inverted-U relationship between the probability that a child is a girl and relative wealth. This relationship is monotonically declining for all other specifications. Source: SICHS census 41 Figure A7: Sex Selection and Relative Wealth (12 largest castes) (a) Ages 0-6 Vanniya Paraiyar Kakliyan Idaiyar Num 0-6 yrs= 35932 Num 0-6 yrs= 23361 Num 0-6 yrs= 8549 Num 0-6 yrs= 4602 .51 .52 .46 .48 .5 .52 .54 .56 .51 .5 .5 .5 .47 .48 .49 .44 .46 .48 .47 .48 .49 0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1 Caste Size = 114047 Caste Size = 65733 Caste Size = 29811 Caste Size = 14557 Labbai Chakkliyan Balija Kammalan Num 0-6 yrs= 4511 Num 0-6 yrs= 1820 Num 0-6 yrs= 1724 Num 0-6 yrs= 1468 .46 .47 .48 .49 .5 .51 .52 .54 .46 .48 .5 .52 .54 .56 .46 .47 .48 .49 .5 .51 .5 .46 .48 0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1 Caste Size = 9773 Caste Size = 4928 Caste Size = 6777 Caste Size = 6159 Boya Naikar Irula Ambattan Num 0-6 yrs= 1331 Num 0-6 yrs= 1230 Num 0-6 yrs= 930 Num 0-6 yrs= 771 .55 .65 .5 .52 .42 .44 .46 .48 .6 .5 .5 .55 .48 .45 .5 .46 .45 .4 0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1 Caste Size = 3448 Caste Size = 3340 Caste Size = 2108 Caste Size = 2449 (b) Ages 7-13 Vanniya Paraiyar Kakliyan Idaiyar Num 7-13 yrs= 38632 Num 7-13 yrs= 26134 Num 7-13 yrs= 10239 Num 7-13 yrs= 4957 .65 .6 .46 .48 .5 .52 .54 .6 .6 .55 .55 .55 .5 .5 .5 .45 .45 .45 0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1 Caste Size = 114047 Caste Size = 65733 Caste Size = 29811 Caste Size = 14557 Labbai Chakkliyan Balija Kammalan Num 7-13 yrs= 4955 Num 7-13 yrs= 2091 Num 7-13 yrs= 1866 Num 7-13 yrs= 1592 .52 .54 .52 .52 .54 .42 .44 .46 .48 .5 .52 .5 .44 .46 .48 .5 .5 .46 .48 .46 .48 0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1 Caste Size = 9773 Caste Size = 4928 Caste Size = 6777 Caste Size = 6159 Boya Naikar Irula Ambattan Num 7-13 yrs= 1526 Num 7-13 yrs= 1200 Num 7-13 yrs= 1001 Num 7-13 yrs= 875 .55 .4 .45 .5 .55 .6 .45 .46 .47 .48 .49 .5 .4 .42 .44 .46 .48 .5 .5 .45 .4 0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1 0 .2 .4 .6 .8 1 Caste Size = 3448 Caste Size = 3340 Caste Size = 2108 Caste Size = 2449 Notes: This figure shows the nonparametric relationship between the probability that the child is a girl and the household’s relative wealth, by caste for children aged 0-6 years in Panel (a) and 7-13 years in Panel (b). The 12 largest castes are shown. Number of 0-6 year old and 7-13 year old children within a caste is mentioned in each individual chart. The caste size refers to the households within the caste. Source: SICHS census 42 Appendix B Omitted Proofs B.1 Proof of Proposition 1 Proof. Equation (6) describes the first order condition associated with the girl’s family’s utility maximization problem: vx vx + vu u′ = 0 ⇒ u′ = − . (15) vu Then the surplus is supermodular and the allocation will be PAM (see Legros and Newman (2007) and Chade et al. (2017)) provided: ∂ 2 v (x, y, u) = vxy + vuy u′ > 0. (16) ∂x∂y From equation (5), we can write the equilibrium utility of the girl’s family as: u−ub 2e 2 u − ub 2(1 − β ) v = log x + y − + + log . (17) β 2 β Next, to derive the condition for PAM in inequality (16), we derive the following terms: 1 vx = u−ub (18) 2e 2 x+y− β u−ub −2 2e 2 vxy = − x + y − (19) β u−ub −e 2 β 1 vu = u−ub + (20) 2e 2 2 x+y− β u−ub −2 u−ub 2e 2 e 2 vuy = − x + y − − . (21) β β 43 Inequality (16) is equivalent to: vx vxy > vuy (22) vu 1 u−ub −2 u−ub  u−ub −2 u−ub  2e 2 2e 2 x+y − β 2e 2 e 2 − x+y− > u−ub − x + y − −  (23) β e 2 β β − β 1 u−ub + 2 x+y − 2e β2 u−ub 1 e 2 −1 > u−ub u−ub (24) −e 2 1 2e 2 β β + 2 x+y− β u−ub e 2 β −1 > u−ub . (25) 1 4e 2 2 x+y− β u−ub 4e 2 If x + y − β < 0 then condition (25) implies that: u−ub 2e 2 x+y− > 0, (26) β which is always satisfied since the LHS is equal to y − d, the consumption of the girl’s parent, which is constrained to be positive. u−ub 4e 2 u − ub β x+y− <0 ⇐⇒ > log (x + y ). (27) β 2 4 The utility from marriage must exceed the outside option of remaining single: x u − ub x mb u ≥ ub + 2 log − mb ⇐⇒ ≥ log − . (28) 2 2 2 2 β mb When y = x, given (28) holds, equation (27) is satisfied provided log 4 (2x) < log x 2 − 2 or equivalently, 2 log β < −mb . Whenever y < x, this sufficient condition is satisfied as well. This establishes the proof. 44 B.2 Proof of Proposition 2 Proof. 1. At the top of the wealth distribution, y = x. When the dowry d = 0, β u(y ) − v (y ) = ub + 2 log y − log ((1 − β )y ) − log(y ) (29) 2 β = ub + 2 log − log(1 − β ). (30) 2 It follows that: β u(y ) − v (y ) > 0 ⇐⇒ ub > log(1 − β ) − 2 log . (31) 2 The right hand side of the second inequality is decreasing in β . Thus, there exists ub (β ), ub ′ (β ) < 0, such that u(y ) − v (y ) > 0 if ub > ub (β ). ub > 0, β ∈ (2/3, 1). For β = 2/3, ub = − log(1/3). For β → 1, ub → 0. Next, we show that u(y ) − v (y ) is increasing in the dowry d: β u(y ) − v (y ) = ub + 2 log (y + d) − log ((1 − β )(y + d)) − log(y − d) (32) 2 ∂ 2 1 1 1 1 u(y ) − v (y ) = − + = + > 0. (33) ∂d y+d y+d y−d y+d y−d Assuming that the condition in Lemma 1 is satisfied, d > 0 and, hence, u(y ) − v (y ) > 0 for any value of the equilibrium dowry. There is sex selection at the top of the wealth distribution. 2. At the bottom of the wealth distribution, girls’ families with wealth y match with boys’ families with wealth x∗ . The last boy to match is indifferent between marrying and staying single: β ∗ x∗ ub + 2 log (x + d) = ub + 2 log − mb . (34) 2 2 Denote Mb ≡ exp(mb /2) > 1. Then we can solve for the equilibrium dowry d(x∗ ) received by the last boy to marry: 1 − Mb β d(x∗ ) = x∗ (35) Mb β Mb β < 1 from Lemma 1 and, hence, d(x∗ ) is increasing in x∗ . Next, we show that u(y ) − v (y ) > 0 for x∗ = x = y . If the family with wealth y = y chooses to have a boy instead of a girl, he will either be unmatched or the last boy to match. Either way, its utility will be 45 y u(y ) = ub + 2 log 2 − mb . y u(y ) − v (y ) = ub + 2 log − mb − log ((1 − β )(x + d(x))) − log y − d(x) (36) 2 y (1 − β )y (2Mb β − 1)y = ub + 2 log − mb − log − log (37) 2 Mb β Mb β (1 − β )(2Mb β − 1) = ub − 2 log 2 − log (38) β2 It follows that: β u(y ) − v (y ) > 0 ⇐⇒ ub > log(1 − β ) − 2 log + log(2Mb β − 1). (39) 2 This is the same condition as at the top of the wealth distribution, except for the log(2Mb β − 1) term. From Lemma 1, Mb β < 1 and, hence, this term is negative. If u(y ) − v (y ) > 0, then u(y ) − v (y ) > 0 for x∗ = x. Next, we show that u(y ) − v (y ) is decreasing in x⋆ for x∗ ≤ x∗ , such that d(x∗ ) = y/2, and increasing in x⋆ for x∗ > x∗ y u(y ) − v (y ) = ub + 2 log − mb − log ((1 − β ) (x∗ + d(x∗ ))) − log y − d(x∗ ) (40) 2 Substituting the expression for d(x∗ ) from equation (35), y (1 − β )x∗ 1 − Mb β u(y ) − v (y ) = ub + 2 log − mb − log − log y − x∗ . (41) 2 Mb β Mb β Therefore: ∂ u(y ) − v (y ) −1 1 1 − Mb β ⋆ = ∗ + (42) ∂x x y− 1−Mb β x∗ Mb β Mb β 1−Mb β −y + 2 Mb β x∗ = (43) 1−Mb β x∗ y − Mb β x∗ 2d(x∗ ) − y = . (44) x∗ y − d(x∗ ) The term in the denominator is positive because y − d(x∗ ), the consumption of the girl’s parent, is constrained to be positive. The term in the numerator is negative if d(x∗ ) ≤ y/2 and positive if d(x∗ ) > y/2. To complete the proof, we thus need to establish that u(y ) − v (y ) > 0 at x∗ such that d(x∗ ) = y/2. From the expression 46 y Mb β for d(x∗ ), the corresponding value of x∗ is 2 1−Mb β . y y Mb β y y u(y ) − v (y ) = ub + 2 log − mb − log (1 − β ) + − log y − (45) 2 2 1 − Mb β 2 2 y 1−β y = ub + 2 log − mb − log − 2 log (46) 2 1 − Mb β 2 1−β = ub − 2 log Mb − log . (47) 1 − Mb β u(y ) − v (y ) > 0 ⇐⇒ ub > log(1 − β ) − log(1 − Mb β ) + 2 log Mb . (48) This condition will be satisfied for Mb → 1 since ub > 0 and, hence, for some mb such that mb < mb . 47 B.3 Proof of Proposition 3 Proof. The extent of sex selection is given by k ⋆ (y ): k ⋆ (y ) = u(y ) − v (µ(y ), y, u(µ(y ))). (49) We need to show that k ⋆ (y ) is increasing in y or ′ k ⋆ (y ) = u′ (y ) − vx µ′ + vy + vu u′ µ′ > 0. (50) From the first order condition (6), along the equilibrium matching µ(y ), it must be that vx + vu u′ = 0, so the derivative can be written as: ′ k ⋆ (y ) = u′ (y ) − (vx + vu u′ )µ′ + vy (51) = u′ (y ) − vy (µ, y, u(µ)). (52) This is increasing provided: −2 1 u(µ(y ))−ub − u(µ(y ))−ub > 0. (53) 4e 2 2e 2 y + µ( y ) − β y + µ( y ) − β To derive the preceding inequality, we note that u′ = − v vx u from the First-Order Condition (15) and that expressions for vx and vu can be obtained from equations (18) and (20). The expression for vy is obtained by partially differentiating expression (5). 1. At the top of the wealth distribution. At y = y , under positive sorting we have y = µ(y ) = x. Then condition (53) can be written as: −2 1 u(y )−ub − u(y )−ub >0 (54) 4e 2 2e 2 2y − β 2y − β β u(y ) = ub + 2 log 2 (y + d) . Substituting the expression for u(y ) in equation (54) and rearranging, we obtain: 1 1 − > 0. (55) d y−d y − d > 0 to satisfy the constraint that the girl’s parent’s consumption must be positive and, hence, the preceding condition will be satisfied if d < y/2. 2. At the bottom of the wealth distribution. At y = y , µ(y ) = x⋆ . As noted above, if the family chooses to have a boy instead of a girl, he will either remain unmatched or be the last boy to match and, hence, y its utility will be u(y ) = ub + 2 log 2 − mb . Therefore u′ (y ) = y 2 . At an income level y , we can then write 48 condition (53), noting that the first term is u′ (y ) from (52) as: 2 1 − u(x∗ )−ub > 0. (56) y 2e 2 x⋆ +y− β Substituting the expression for u(x∗ ) in equation (56) and rearranging, we obtain: 2 1 − > 0. (57) y y − d(x∗ ) y − d(x∗ ) > 0 to satisfy the constraint that the girl’s parent’s consumption must be positive and, hence, it is straightforward to verify that the preceding condition will be satisfied if d(x∗ ) < y/2. 49 References Alfano, M. (2017): “Daughters, dowries, deliveries: The effect of marital payments on fertility choices in India,” Journal of Development Economics, 125, 89–104. Ambrus, A., E. Field, and M. Torero (2010): “Muslim family law, prenuptial agreements, and the emergence of dowry in Bangladesh,” The Quarterly Journal of Economics, 125, 1349–1397. Anderson, S. (2003): “Why Dowry Payments Declined with Modernization in Europe but are Rising in India,” Journal of Political Economy, 111, 269–310. ——— (2007): “The Economics of Dowry and Brideprice,” The Journal of Economic Perspectives, 21, 151– 174. Anderson, S. and C. Bidner (2015): “Property Rights over Marital Transfers,” The Quarterly Journal of Economics, 130, 1421–1484. Anukriti, S. (2018): “Financial incentives and the fertility-sex ratio trade-off,” American Economic Journal: Applied Economics, 10, 27–57. Anukriti, S., S. Kwon, and N. Prakash (2022): “Saving for dowry: Evidence from rural India,” Journal of Development Economics, 154, 102750. Arnold, F., M. K. Choe, and T. K. Roy (1998): “Son Preference, the Family-Building Process and Child Mortality in India,” Population Studies, 52, 301–315. Arnold, F., S. Kishor, and T. Roy (2002): “Sex-Selective Abortions in India,” Population and Devel- opment Review, 28, 759–785. Arokiasamy, P. and S. Goli (2012): “Explaining the skewed child sex ratio in rural India: Revisiting the landholding-patriarchy hypothesis,” Economic and Political Weekly, 85–94. Ashraf, N., N. Bau, N. Nunn, and A. Voena (2020): “Bride price and female education,” Journal of Political Economy, 128, 591–641. Attanasio, O. P. and G. Weber (1993): “Consumption Growth, the Interest Rate and Aggregation,” The Review of Economic Studies, 60, 631–649. Basu, A. M. (1999): “Fertility Decline and Increasing Gender Imbalance in India, Including a Possible South Indian Turnaround,” Development and Change, 30, 237–263. Bau, N., G. Khanna, C. Low, and A. Voena (2023): “Traditional Institutions in Modern Times: Dowries as Pensions When Sons Migrate,” Tech. rep., National Bureau of Economic Research. Becker, G. S. (1973): “A theory of Marriage: Part I,” Journal of Political Economy, 81, 813–846. 50 Bhalotra, S., A. Chakravarty, and S. Gulesci (2020): “The price of gold: Dowry and death in India,” Journal of Development Economics, 143, 102413. Bhalotra, S., A. Chakravarty, D. Mookherjee, and F. J. Pino (2019): “Property rights and gender bias: Evidence from land reform in West Bengal,” American Economic Journal: Applied Economics, 11, 205–37. Bhalotra, S. and T. Cochrane (2010): “Where Have All the Young Girls Gone? Identification of Sex Selection in India,” IZA Discussion Paper No. 5381. Bhaskar, V. (2011): “Sex Selection and Gender Balance,” American Economic Journal: Microeconomics, 3, 252–253. Bhat, P. M. and S. S. Halli (1999): “Demography of Brideprice and Dowry: Causes and Consequences of the Indian Marriage Squeeze,” Population Studies, 53, 129–148. Bloch, F. and V. Rao (2002): “Terror as a bargaining instrument: A case study of dowry violence in rural India,” American Economic Review, 92, 1029–1043. Blundell, R., M. Browning, and C. Meghir (1994): “Consumer Demand and the Life-Cycle Allocation of Household Expenditures,” The Review of Economic Studies, 61, 57–80. Botticini, M. (1999): “A Loveless Economy? Intergenerational Altruism and the Marriage Market in a Tuscan Town, 1415–1436,” The Journal of Economic History, 59, 104–121. Botticini, M. and A. Siow (2003): “Why Dowries?” American Economic Review, 93, 1385–1398. Caldwell, J. C., P. H. Reddy, and P. Caldwell (1983): “The Causes of Marriage Change in South India,” Population Studies, 37, 343–361. Calvi, R., A. Keskar, et al. (2021): “’Til Dowry Do Us Part: Bargaining and Violence in Indian Families,” Tech. rep., CEPR Discussion Papers. Chade, H., J. Eeckhout, and L. Smith (2017): “Sorting through Search and Matching Models in Economics,” Journal of Economic Literature, 55, 493–544. Chiplunkar, G. and J. Weaver (2023): “Marriage markets and the rise of dowry in India,” Journal of Development Economics, 164, 103–115. Choo, E. and A. Siow (2006): “Who Marries Whom and Why,” Journal of Political Economy, 114, 175–201. Corno, L., N. Hildebrandt, and A. Voena (2020): “Age of marriage, weather shocks, and the direction of marriage payments,” Econometrica, 88, 879–915. Das, M., W. K. Newey, and F. Vella (2003): “Nonparametric estimation of sample selection models,” The Review of Economic Studies, 70, 33–58. 51 Desai, S. and L. Andrist (2010): “Gender scripts and age at marriage in India,” Demography, 47, 667–687. Doepke, M. and M. Tertilt (2009): “Women’s Liberation: What’s in it for Men?” The Quarterly Journal of Economics, 124, 1541–1591. Dyson, T. and M. Moore (1983): “On Kinship Structure, Female Autonomy, and Demographic Behavior in India,” Population and Development Review, 35–60. Edlund, L. (1999): “Son Preference, Sex Ratios, and Marriage Patterns,” Journal of Political Economy, 107, 1275–1304. Field, E. and A. Ambrus (2008): “Early Marriage, Age of Menarche, and Female Schooling Attainment in Bangladesh,” Journal of Political Economy, 116, 881–930. Field, E., R. Pande, N. Rigol, S. Schaner, and C. Troyer Moore (2021): “On Her Own Account: How Strengthening Women’s Financial Control Impacts Labor Supply and Gender Norms,” American Economic Review, 111, 2342–75. Foster, A. and M. Rosenzweig (2001): “Missing Women, the Marriage Market and Economic Growth,” Brown mimeo. Geddes, R. and D. Lueck (2002): “The Gains from Self-Ownership and the Expansion of Women’s Rights,” The American Economic Review, 92, 1079–1092. George, S., R. Abel, and B. D. Miller (1992): “Female Infanticide in Rural South India,” Economic and Political Weekly, 1153–1156. Greenwood, J., N. Guner, and G. Vandenbroucke (2017): “Family Economics Writ Large,” Journal of Economic Literature, 55, 1346–1434. Guilmoto, C. Z. (2009): “The Sex Ratio Transition in Asia,” Population and Development Review, 35, 519–549. Gupta, B. (2014): “Where have all the brides gone? Son preference and marriage in India over the twentieth century,” The Economic History Review, 67, 1–24. Heckman, J. and S. Navarro-Lozano (2004): “Using matching, instrumental variables, and control functions to estimate economic choice models,” Review of Economics and Statistics, 86, 30–57. Heckman, J. J. (1976): “Simultaneous equation models with both continuous and discrete endogenous variables with and without structural shift in the equations,” Studies in Nonlinear Estimation, Ballinger. Hussam, R. N. (2021): “Sex Selection and the Indian Marriage Market,” . Jayachandran, S. (2017): “Fertility Decline and Missing Women,” American Economic Journal: Applied Economics, 9, 118–139. 52 Jayachandran, S. and I. Kuziemko (2011): “Why do mothers breastfeed girls less than boys? Evidence and implications for child health in India,” The Quarterly Journal of Economics, 126, 1485–1538. Jayachandran, S. and R. Pande (2017): “Why are Indian children so short? The role of birth order and son preference,” American Economic Review, 107, 2600–2629. Jeffery, R., P. Jeffery, and A. Lyon (1984): “Female Infanticide and Amniocentesis,” Social Science & Medicine, 19, 1207–1212. Jejeebhoy, S. J. and Z. A. Sathar (2001): “Women’s Autonomy in India and Pakistan: the Influence of Religion and Region,” Population and Development Review, 27, 687–712. Kabeer, N., L. Huq, and S. Mahmud (2014): “Diverging stories of “missing women” in South Asia: Is son preference weakening in Bangladesh?” Feminist Economics, 20, 138–163. Kapadia, K. (1993): “Marrying Money: Changing Preference and Practice in Tamil Marriage,” Contribu- tions to Indian Sociology, 27, 25–51. ——— (1995): Siva and her Sisters: Gender, Caste, and Class in Rural South India., Westview Press, Inc. Legros, P. and A. F. Newman (2007): “Beauty is a Beast, Frog is a Prince: Assortative Matching with Nontransferabilities,” Econometrica, 75, 1073–1102. Lindenlaub, I. (2017): “Sorting multidimensional types: Theory and application,” Review of Economic Studies, 84, 718–789. Moorjani, P., K. Thangaraj, N. Patterson, M. Lipson, P.-R. Loh, P. Govindaraj, B. Berger, D. Reich, and L. Singh (2013): “Genetic Evidence for Recent Population Mixture in India,” The American Journal of Human Genetics, 93, 422–438. Munshi, K. and M. Rosenzweig (2016): “Networks and Misallocation: Insurance, Migration, and the Rural-Urban Wage Gap,” The American Economic Review, 106, 46–98. `ze (1995): “Mortality, fertility, and gender bias in India: A district- Murthi, M., A.-C. Guio, and J. Dre level analysis,” Population and Development Review, 745–782. Neelakantan, U. and M. Tertilt (2008): “A Note on Marriage Market Clearing,” Economics Letters, 101, 103–105. Prasad, B. D. (1994): “Dowry-related violence: A content analysis of news in selected newspapers,” Journal of Comparative Family Studies, 71–89. Rahman, L. and V. Rao (2004): “The Determinants of Gender Equity in India: Examining Dyson and Moore’s Thesis with New Data,” Population and Development Review, 30, 239–268. Rao, V. (1993a): “Dowry ‘Inflation’in Rural India: A Statistical Investigation,” Population Studies, 47, 283–293. 53 ——— (1993b): “The Rising Price of Husbands: a Hedonic Analysis of Dowry Increases in Rural India,” Journal of Political Economy, 101, 666–677. Robinson, P. M. (1988): “Root-N-consistent semiparametric regression,” Econometrica, 931–954. Rosenzweig, M. R. and T. P. Schultz (1982): “Market Opportunities, Genetic Endowments, and In- trafamily Resource Distribution: Child Survival in Rural India,” The American Economic Review, 72, 803–815. Sekher, T. (2010): “Special Financial Incentive Schemes for the Girl Child in India: a Review of Select Schemes,” United Nations Population Fund. Sen, A. (1990): “More than 100 Million Women are Missing,” The New York Review of Books. Sharma, U. (1980): Women, Work and Property in North-West India, Routledge Kegan & Paul. Srinivas, M. N. (1984): Some Reflections on Dowry, Published for the Centre for Women’s Development Studies, New Delhi, by Oxford University Press. ——— (1989): The Cohesive Role of Sanskritization and other essays, Oxford University Press, USA. Tertilt, M. (2005): “Polygyny, Fertility, and Savings,” Journal of Political Economy, 113, 1341–1371. Zhang, J. and W. Chan (1999): “Dowry and Wife’s Welfare: A Theotrical and Empirical Analysis,” Journal of Political Economy, 107, 786–808. 54