Female Early Marriage and Son Preference in Pakistan

Abstract In this study, we employ pooled data from four rounds of Pakistan Demographic and Health Survey (PDHS) to examine whether, and to what extent, does the incidence of early marriage shape the married women’s perspectives on gender preference associated with reproduction. We employ a number of econometric techniques (Probit, OLS, Cox Hazard Model, IV Probit and treatment effects) and a large set of model specifications, and find significant evidence supporting the role of early marriage in perpetuating disproportionate preference for boys. Women who married before turning 18 not only state a greater desire for boys but are also less likely to stop reproduction as long as they do not have a boy. Early-age marriage is associated with 7.7–12.5 per cent higher incidence of fertility discontinuation among women without a son. This son-preferring behaviour is stronger at higher birth order and also reflects in differential spacing patterns. Women’s education appears to be the strongest channel through which these effects are mediated. The divergence between early- and late-marrying women appears to have sharpened over time. The findings of this study underscore the role played by early marriage in altering the gender-specific attitudes prevalent in the society, and highlight existing gender inequality traps.


Introduction
Thirty-nine per cent of Pakistani women of child-bearing age are reported to get married before reaching the age of 18 (PDHS 2017(PDHS -2018. Though lower than 50 per cent reported in 1990 (PDHS 1990(PDHS -1991, this incidence still remains high by world average. The practice of early-age marriage, also called child marriage, results from a higher perceived value of the young bride. In traditional societies, younger women are believed to be more fertile, sexually inexperienced and easy to 'control'. For parents, early marriage of daughters implies lower spending on education, less effort chaperoning the girl in order to protect virginity and guard 'family honour', and smaller dowry requirement (Allendorf, Thornton, Mitchell, Young-DeMarco, & Ghimire, 2017).
Early marriage has important consequences for the health and well-being of the mother and the child. Women who marry early produce more children than who marry later (Maitra, 2004;Nasrullah, Muazzam, Bhutta, & Raj, 2014;Raj, Saggurti, Balaiah, & Silverman, 2009). They are younger at the time of first birth and have subsequent births at shorter intervals (Jensen & Thornton, 2003;Koski, Clark, & Nandi, 2017;Raj, 2010). Early marriage is associated with greater risk of still birth and miscarriages (Kamal & Hassan, 2015). There is increasing evidence for adverse health outcomes among children born to women who married at an early age, including higher risk of premature birth, neo-natal, infant, or child mortality (Adhikari, Bott, Jejeebhoy, Shah, & Puri, 2003;Garcia-Hombrados, 2017;Raj et al., 2010) as well as negative effects on child weight, height, and general health (Chari, Heath, Maertens, & Fatima, 2017;Palloni, 2017;Wachs, 2008). Women with early marriages show higher psychological distress, whereas women with late marriages tend to adjust better (Shaud & Asad, 2020). Early marriage can also limit women's economic empowerment and education outcomes of their children (Sekhri & Debnath, 2014;Yount, Crandall, & Cheong, 2018).
A related area of investigation pertains to the influence of female early marriage in sustaining prevailing gender bias in general, and gender-specific reproductive outcomes in particular. Asadullah and Wahhaj (2019) find that early marriage increases agreement with statements supportive of traditional gender roles and gender bias in the allocation of resources. They hypothesize four potential pathways through which female early marriage in developing countries can affect women's beliefs and attitudes towards traditional gender norms: (1) less schooling and exposure to a school curriculum presenting alternative views; (2) fewer social networks; (3) lower likelihood of matching with more progressive men; and (4) earlier experience of marital responsibilities.
The gender perspectives of early-married women get shaped by their degree of empowerment in important issues, such as contraceptive use and reproductive choices (Larsson & Stanfors, 2014;Upadhyay & Karasek, 2012).
In this study, we employ pooled data from four Pakistan Demographic and Health Surveys (PDHS) to examine whether, and to what extent, does the incidence of early marriage shape the married women's perspectives on gender preference associated with reproduction. We investigate how early-age marriage influences both the reported or revealed son preference of the mother (observed in differential stopping) as well as the stated or desired preference that the interviewed women state. From a medical point of view, the likelihood of bearing sons does not depend on the age a woman marries. Any significant variation in the number of sons she bears should reflect differential fertility stopping patterns ultimately resulting from differences in the socioeconomic profile of early and late-marrying women. 1 To our knowledge, this is the first study that addresses this subject. We employ a number of econometric techniques (Probit, OLS, Cox Hazard Model, IV Probit, and treatment effects) and a large set of model specifications to support our analysis. We find significant evidence supporting the role of female early marriage in perpetuating disproportionate preference for boys in Pakistan. Women who married before turning 18 not only state a greater desire for boys but are also more likely to end reproduction only after obtaining the desired number of boys. This son-preferring behaviour is stronger at higher birth order and also visible in differential spacing patterns. The divergent trend seems to be stronger in the post-2000 cohort compared to the women who married before 2000, reflecting increasing social pressures associated with demographic transition. Women's schooling appears to be the most important channel through which these gender-specific reproductive effects are mediated. We also find evidence for reported son preference among early-married men. These findings are robust to the use of alternative definitions and empirical procedures.
Our study is organized as follows: Section 2 presents the survey data and shows relevant salient statistics. Section 3 describes the empirical model and the outcome and control variables employed in the estimations. Section 4 reports key findings followed by robustness checks and additional results in Section 5. Section 6 concludes.

Data
We pool data of all the four rounds (1990-1991, 2006-2007, 2012-2013, and 2017-2018) of the Pakistan Demographic and Health Survey (PDHS). The PDHS are household surveys representative at the national level, containing information about fertility, family planning, maternal and child health. A two-stage stratified sample design was adopted for the survey. The pooled sample consists of 45,260 women who married between 1951 and 2018. Out of these, 21,849 women are considered to have completed their fertility. The latter group corresponds to the women who gave the answer 'want no more children' in response to the question 'Do you desire more children?', those who report to be infecund or who they or their husbands had undergone sterilization procedure. In our sample, 17,528 women reported to desire no more children, 1348 reported to be infecund, whereas 2973 were sterilized. According to the dataset, 43 per cent of the women of child-bearing age interviewed got married before the age of 18 while 24 per cent gave birth to their first child before turning 20. The data show substantial difference between the profile of women who married before the age of 18 and those who married later (Table 1). Fewer early-marrying women and their husbands went to school than did their later-marrying counterparts. On average, early marrying women are poorer (46% of early-marrying women vs. 30% of late-marrying women) and less urban (28% early-marrying women vs. 39% latemarrying women). Besides, a greater proportion of them work for a living than do women who married later. Table 2 sheds light on divergent reproductive behaviours of the two groups of women by comparing the actual and desired number of sons born to early-and late-marrying women. On average, early-marrying women are found to have more boys (3.21) than late-marrying women (2.51). The difference in the mean number of sons is clearer at higher parities. Similarly, women who marry early state greater desire for boys (2.29 boys) compared with those who marry later (2.00 boys). The spacing pattern of early-marrying women too differs from that of their latemarrying cohorts. Except for the interval between the first and the second birth, the spacing for all intervals is higher among early-marrying women.

Empirical framework
We estimate early marriage's association with son preference by regressing indicators of women's revealed and stated son preference on the early marriage indicator and controlling for Source: Authors' calculations using pooled data from the four rounds of PDHS. The means are reported in columns 1 and 2. Columns 3 report the t-statistic for the early marriage-late marriage mean comparison test.
Female early marriage and son preference in Pakistan 1551 individual and household socioeconomic factors. The empirical model for son preference can be given as follows: Where SP ij is an indicator of the son preference of the woman i belonging to the household j, EM ij represents woman's early marriage, that is, whether the woman i belonging to the household j married before reaching the age of 18, X ij represents individual characteristics of the woman i belonging to the household j: Y J represents the characteristics of household j, and l ij is the error term.
The variable of interest is woman's early marriage which is a binary variable that takes the value of 1 if the woman gets married before age 18, 0 otherwise. Article 1 of the Convention of the Rights of the Child (CRC) defines child or early marriage as legal or customary marriage between two people, of whom one or both spouses are below the age of 18. Later, we obtain our robustness estimates by using the count variable for women's age at marriage.
Individual and household characteristics include the respondent woman's age, education level, employment status, exposure to electronic or print media, age difference with the husband, husband's education level, household size, household wealth, area of residence (urban/rural) and access to improved water supply. The household wealth variable is constructed by generating an index of household assets, such as home ownership, floor type, water source, electricity availability and durable consumer goods. The quintiles of the variable generated using Principal Component Analysis indicate the economic status of the household ranging from the poorest to the richest quintile. Table A1 provides the definitions of the variables included and their means and proportions. About 43 per cent of the women reported to have married before age 18 compared to 10 per cent of the husbands. The mean age of women is 32 years. The average age difference between husband and wife were 5.51 years. Majority of women in our sample possessed no formal education (60%) compared to 35 per cent of the husbands. Likewise, only 8 per cent women report having acquired tertiary level education compared to 15 per cent husbands, respectively. Around one-fourth (22%) of women in our sample participate in the labour market. Forty-one per cent of women report either listening to radio or watching television at least once a week. The average household size is 8.26. About two-thirds of the households (66%) live in rural areas. Source: Authors' calculations using pooled data from the four rounds of PDHS. The means are reported in columns 1 and 2. Columns 3 report the t-statistic for the early marriage-late marriage mean comparison test.
3.1. Outcome variables 3.1.1. Revealed son preference. We employ three indicators to represent different dimensions of revealed son preference. Following Javed and Mughal (2022), we define the baseline indicator of differential birth stopping for revealed or reported son preference as a binary variable which takes the value of 1 if a woman has no son at a given birth order and does not pursue further child birth, 0 otherwise. The second variable modifies this definition to focus only on the sex of the latest child. For example, the variable takes the value of 1 if a woman who has four children, out of which the fourth is a girl, does not proceed to subsequent child birth. The third indicator of revealed son preference is a count variable that pertains to succeeding birth spacing. It is defined as the succeeding birth space in months at a given parity if a woman has no son. Estimation results of these three outcomes are reported for the first four birth orders.
3.1.2. Stated son preference. We employ two indicators to measure stated or desired son preference: First, following Behrman and Duvisac (2017), we define the baseline indicator for stated or desired son preference as a binary variable which takes the value of 1 if the woman's desire number of sons exceeds the desired number of daughters. Following Gaudin (2011), we also use a ratio to denote desired son preference. The alternative measure is defined as the ratio of the difference between Ideal number of boys and girls to the Ideal number of children.
In our sample, 14, 16, 18, and 20 per cent woman without at least one son stop their childbearing at the first, second, third, and fourth birth order, respectively. The mean succeeding birth space of women without at least one son ranges from 27.31 to 28.63 months at the first four birth orders. The average spacing is the shortest at the fourth birth order. Thirty-seven per cent women reported desiring more boys than girls. Figure 1 represents the gender-specific progression to subsequent birth. Almost all (98%) of the women moved on to second parity regardless of whether the firstborn was a boy or a girl (Figure 1). At higher parities however, the reproductive patterns of girls-only women diverge from those of women with one or more sons: 96 per cent of women with no boys proceed to third birth compared to 91 per cent of women with one or two sons, while 93 per cent of girls-only women move on to fourth birth compared to 83 per cent of women with three sons.

Methodology
Given the binary nature of the first two revealed preference outcomes and the first stated preference outcome, the corresponding models are initially estimated using the Probit estimator,  Female early marriage and son preference in Pakistan 1553 while the stated preference ratio is estimated using OLS. Later, these estimations are carried out using instrumental-variable and matching estimators. The spacing outcome is regressed using the Cox Hazard Model. All estimations are carried out first without, and then with the full set of controls and region-and time-fixed effects.

Revealed son preference
In son preferring societies, discriminatory reproductive patterns manifest themselves, either in the form of sex-selective abortions, or differential birth stopping. Even though preferential attitudes towards boys are widespread in the Indian Subcontinent, the practice does not usually enjoy good press (Robitaille, 2013). Islam, Pakistan's dominant religion, does not promote sexselective reproductive practices (Aydede & Robitaille, 2019). There is little evidence supporting widespread practice of sex-selective abortion in Pakistan (see for instance Zaidi & Morgan, 2016). However, differential stopping is reported to be widely practised (Hussain, Fikree, & Berendes, 2000;Javed & Mughal, 2022). In the presence of disproportionate preference for male offspring, early-marrying women continue childbearing as long as the desired number of sons is not attained. Table 3 reports partial results for the association between women's early marriage and their child-stopping behaviour in the situation where all the existing children are girls. We report both the estimates and the marginal effects (ME). Columns 1-8 alternately show results of Probit estimations, with and without controls, and region-and time-fixed effects, for the likelihood of childbearing after the first, second, third, and fourth birth, respectively. The results are negative and statistically significant at all the birth orders. At the first birth order, an early-marrying woman with a girl child is 7.7-10.5 per cent less likely to stop child bearing, in contrast to the late-marrying women. The impact is found to be stronger at higher birth orders. At the second birth order, the presence of no son is associated with 9.2 per cent (without controls) and 12.4 per cent (with controls) lower probability of stopping childbearing. In other words, early-marrying women, both of whose first two children are girls, are 9.2-12.4 per cent less likely to discontinue fertility compared with late-marrying women. The corresponding results for probit estimates for birth order 3 and 4 show 8.3-12.5 per cent lower likelihood of stopping childbearing among early-marrying women without at least one son.
This differential birth stopping effect is similar for both rural and urban women, 2 except for the fourth birth on which the effect is much stronger among urban women (17.2% less likelihood) compared to the rural women (9% less likelihood). The proportion of early-marrying women living in the urban areas (0.28) is much lower than that of the late-marrying women (0.39). This, combined with the fact that early-marrying women have on average more children and a greater stated desire for sons, makes the revealed son preference effect of early marriage stronger in the cities.
These results could be challenged on the grounds that the fertility preferences of early-marrying women might be over-represented in the sample. At a given time, women who marry early are more likely to achieve their desired fertility than women who married late and thereby began their reproductive phase later. 3 We estimate our baseline model on the sub-sample of women whose fertility could be considered complete, that is, women who gave the answer 'want no more children' in response to the question 'Do you desire more children?', those who report to be infecund, or who they or their husbands had undergone sterilization procedure. The association between early marriage and reported preference remains negative and statistically significant as before. However, the marginal effects are much lower (1.3-6.7%) compared to those observed in the full sample.
The definition of 'completed fertility' used in our above estimations for the subsample with complete fertility is based on the interviewed women's self-reported state of infecundity. Now we employ an alternative definition of complete fertility by restricting our sample to women age Female early marriage and son preference in Pakistan 1555 40 or above. One can assume that by that age, most women have completed their fertility, and many of those who have not are nearing menopause and are facing difficulty conceiving. As before, the association between early marriage and revealed son preference remains negative and significant. The marginal effects range between 1.2 and 6.0 per cent. Table 4 presents the estimates of childbearing behaviour with respect to the sex of the last child. Compared to late-marrying women, early-marrying women are more likely to continue childbearing if the last child happens to be a girl with the marginal effects ranging from 7.7 to 13 per cent.
Another dimension of fertility, but for which we do not find much evidence, is the differential child-spacing practised by early-married Pakistani women. Javed and Mughal (2020) report strong evidence for differential behaviour at early parities. They find that women whose first or second child is a son have significantly longer subsequent birth intervals compared with women with no sons. In this study, our interest lies not in the spacing patterns of women with one or more sons per se, but rather in their interaction with early marriage. Table 5 reports partial results for Cox Hazard Model estimations for the first four parities. We observe significant difference between early-and late-marrying women at the first two birth orders. Early-marrying women without a son have 6.7-8.4 per cent shorter subsequent birth interval in contrast to late marrying woman. The effect is not visible at higher birth orders.
All in all, these findings show a clear difference between women who married early and those who married later in terms of their revealed son preference reflected in greater incidence of differential birth stopping and spacing.

Stated son preference
Next, we examine if the divergence in preference for male child found between women who married early and those who married later is also reflected in their stated desire. Stated preference, to some extent, reflects the woman's perception of gender equality, and should plausibly decrease with growing maturity and autonomy that accompanies later marriage. Partial results of Probit estimations for stated son preference reported in Table 6 support this argument. The results shown in Columns 1 and 2 point to a positive and mostly significant relationship between desired son preference and early marriage. The marginal effects are similar to those observed for revealed preference, and range from 1.8 to 6.3 per cent.
Here, the survey questions on which the stated preference variable used in the above set of estimations is based merit scrutiny. Women who already given birth to a child were asked the following questions to find out their desired fertility preferences: "If you could go back to the time you did not have any children and could choose exactly the number of children to have in your whole life, how many would that be?", "How many of these children would you like to be boys?" and "How many would you like to be girls, and for how many would it not matter if it's a boy or a girl?" Responses to such questions, constituting direct measures of son preference, are criticized in the literature for being subject to Rationalization bias. A woman's perception of ideal number of sons and daughters may be driven by the number of sons she has already borne (Dasgupta, 2016;Pritchett, 1994). In our data however, we do not find support for this assertion. The correlation between the indicators of stated and revealed preference is low (correlation coefficient ¼ À0.0382 for parity 1, À0.06 for parity 2, À0.04 for parity 3, and À0.08 for parity 4). Furthermore, the results of the stated son preference model are not much affected if the sex of existing children is controlled for. For this, we include the son ratio variable, defined as the ratio of sons as a proportion of the total number of children born to the woman. The coefficient of the variable is found to be significant, and its inclusion, if anything, improves the statistical significance of the models. An early-marrying woman with completed fertility is Female early marriage and son preference in Pakistan 1557 1.8-6.3 per cent more likely to declare greater desire for sons than daughters than a later-marrying woman does (Columns 3-4).
Next, we employ another definition of stated son preference, defined as the ratio of the difference between Ideal number of boys and girls to the Ideal number of children born to the woman. Partial results for OLS estimations carried out with this alternative measure (shown in columns 5-8) again point to a positive and statistically significant association between female early marriage and stated son preference. The coefficients of the early-marriage variable for the models with or without controls or fixed effects, or with or without the inclusion of the sex of existing children, are all significant and lie in the 1-3.5 per cent range.
The results for women residing in rural and urban areas are quite similar, and suggest that the differences between the early-and late-marrying women's stated preference effects do not substantially differ by their place of residence.

Husband's early marriage
Next, we compare the gender-specific fertility effects of female early marriage examined thus far with those associated with men's early marriage. Early marriage is much less common among Pakistani men. Only 10 per cent of men in our sample got married before age 18 compared to 43 per cent of women. This notwithstanding, male early marriage's association with revealed son preference does not appear much different from that observed in the case of female early marriage. Table 7 reports results of husband's early marriage and revealed and stated son preference. The marginal effects for the four parities range from 5.6 to 7.6 per cent (columns 1-4). This finding is in line with the observation that historically, son preference and demand for additional children has been strong both among Pakistani men and women (Khan & Sirageldin, 1977).
Interestingly, the association of men's early marriage with stated son preference is insignificant (columns 5-8), suggesting that unlike women, men's stated desire for boys does not significantly differ by age at marriage. In other words, while early-marrying women state a desire for boys that is significantly greater than that expressed by late-marrying women, early-and late-marrying men exhibit no such difference in their stated gender preference. Compared to men, women in patriarchal societies face more pressure to produce sons (Javed & Mughal, 2019) which is expressed in their greater stated desire for boys. Source: Authors' calculations using pooled data from the four rounds of PDHS. Columns 1-4 present results for the subsequent birth space at the nth birth order with the set of controls. Controls include woman's characteristics (age, age difference with husband, education, employment status, media exposure), spouse education, household size, wealth status, and geographical features (place of residence, region). The sample is restricted to women without a son at the nth birth order. Robust standard errors in parentheses. ÃÃÃ p < 0.01, ÃÃ p < 0.05, Ã p < 0.1. Female early marriage and son preference in Pakistan 1559

Demographic transition
Women in our sample got married between 1951 and 2018. During this time period, Pakistan went through demographic transition, with fertility rates falling from over 6 in the 1950s to <4 in 2018 (World Bank, 2020). Contraceptive prevalence increased from 12 per cent in 1991 to 34 per cent in 2018 (PDHS, 2018). In the presence of son preferring norms, smaller family requirements can aggravate gender-specific fertility stopping. The change can be expected to affect early-marrying women disproportionately, leading to increasing difference with late-marrying women. We find evidence for this argument by comparing the pre-and post-2000 marriage cohorts. Table 8 reports partial results of probit estimations for these two cohorts, without and with the set of controls and fixed effects. The association between early marriage and reported son preference remains negative and statistically significant. The effect is substantially stronger in post-2000 marriages. The likelihood of stopping childbearing without a son among the more recent (post-2000) early-marrying women is lower by as much as 21 per cent (parity 2/3) compared to their later-marrying counterparts.

Mediating channels
We examine the role of the four mediating channels suggested by Asadullah and Wahhaj (2019) through which early marriage can affect women's beliefs and attitudes towards traditional gender norms.
(1) Women's schooling: we compare women with no schooling to those with at least some education.
(2) Social network: We compare women who, in response to the question: 'Who usually decides on visits to family or relatives?', answer: 'respondent alone' or 'Respondent and husband/partner jointly', with those who answer: 'family elders', 'husband/partner alone', or 'others'. (3) Progressive spousal matching: We compare women whose husbands have acquired at least five years or more education with those who do not. (4) Earlier experience of marital responsibilities: We compare women who gave birth to their first child within twelve months of marriage with those who did not.
Out of the four channels, women's education appears to be the strongest, especially considering the women's revealed son preference. 4 The education profiles of the early-and late-marrying women are substantially different. Only 26 per cent of early-marrying women have ever been to school compared to 47 per cent of late-marrying women. Early-marrying women, on average, hold 1.68 years of schooling compared to 4.2 years for late-marrying women. The differences in gender-specific birth stopping effects between the early-and late-marrying women are stronger among educated women, particularly at higher birth orders. For example, at the third birth order, illiterate early-marrying women without a son are 8.6 per cent less likely to stop fertility compared to late-marrying women, while their educated counterparts are 18.2 per cent less likely to stop childbearing compared to corresponding late-marrying women. The difference is much less important in case of stated preference.
The difference in gender-specific birth stopping behaviour of early-and late-marrying women is weaker for the other three mediation channels examined. The difference among women with weak social network ranges between 12.8 and 13.9 per cent at the four birth orders, and 7.6 and 13.6 per cent among women with stronger social network. Both the revealed and stated preference effects are somewhat stronger among women with weak network. Likewise, the effects are stronger among women who take up marital responsibilities soon after marriage. In contrast, the differential child stopping impact of early marriage is similar across women with and without progressive spousal matches.

Robustness measures and additional results
In this section, we show that the main results are robust to a wide range of robustness checks. 5 Female early marriage and son preference in Pakistan 1561

Source:
Authors' calculations using pooled data from the four rounds of PDHS. Columns 1-8 present results for the subsequent birth at the nth birth order, with the set of controls and fixed effects. Controls include woman's characteristics (age, age difference with husband, education, employment status, media exposure), spouse education, household size, wealth status, and geographical features (place of residence, region). The sample is restricted to women without a son at the nth birth order. Robust standard errors in parentheses.

Instrumental variable estimations
The estimations reported in the previous section may be subject to endogeneity concerns. Personal and family values, local traditions and cultural norms prevalent in the society influence the age at which girls get married and the importance the birth of boys enjoys. Age at which a woman marries, therefore, cannot be treated as a random event, and may plausibly be correlated with unobserved factors which also affect the woman's reproductive preferences. We employ an instrumental strategy to tackle potential endogeneity present in our estimated models. We construct a community-level instrument for this purpose. The instrument is defined as the percentage of incidence of early marriage observed in the Primary Sampling Unit (PSU) among the women who married before the respondent. This instrument takes its inspiration from Delprato, Akyeampong, and Dunne (2017), who analyse the inter-generational education effects of early marriage in sub-Saharan Africa. The logic for this instrument goes as follows: In traditional societies, marriage is an institution meant to maintain and promote social ties and within-group connections (UNFPA, 2006). Early-age marriage, in this context, serves as a means to preserve community values related to gender and sexuality. The community is expected to adhere to the practice, and failure to conform could be socially costly (Bayisenge, 2010;Bicchieri & Mercier, 2014;Srinivas, 2000). The proportion of early-married women in the community could therefore act as a useful community-level predictor of female age at marriage, while not being directly related to individual reproductive outcomes. The first-stage regressions of the instrumental-variable estimation confirm this plausible association. The communityincidence variable is strongly associated with the female age at marriage variable, with all coefficients significant at the 1 per cent level. The F-statistics is above 10 across all specifications, implying that instrument is strong. The results of the IV Probit estimations for fertility stopping show that the impact of female early marriage on fertility stopping among women without a son at a given birth order is negative and significant as before, with coefficients ranging between À1.06 and À1.78. The impact of female early marriage on the stated desire for sons is likewise negative and statistically significant, with coefficients ranging from 0.1 to 0.6. We compute the proportion of early-marrying women among all the women in the PSU who got married prior to the surveyed woman. The trends and norms prevalent at the time of the surveyed woman's marriage might closely relate to those present at the time of marriages that took place in the near past. The instrument may in such a case not be considered exogenous. We consider this possibility by limiting the marriages in the PSU to five and ten years prior to the respondent's marriage, respectively. The results of IV Probit estimations for revealed and stated son preference carried out with these two instruments are similar to those found previously, and again point to a lower likelihood of birth stopping among early-marrying women without a son.

Treatment effects
It is possible that women who marry early self-select based on their individual and household characteristics. As previously shown in Table 1, early and late-marrying women differ substantially on a number of observables including schooling, employment, wealth status and place of residence. Women marrying early may therefore differ from those marrying later in ways that could be considered non-random. We account for this possibility of selection bias by estimating the baseline model using different treatment effect estimations including Propensity Score Matching (PSM), Regression Adjustment (RA), Inverse probability weights (IPW) and Augmented IPW (AIPW). The PSM is a matching technique which matches the treated group individuals (those who married early) to the non-treated counterparts (women who married later) based on a propensity score for participation given observable characteristics of the individual. RA estimates the average treatment effect (ATE) and the potential-outcome means (POMs) from observational data. RA estimators use contrasts of averages of treatment-specific Female early marriage and son preference in Pakistan 1563 predicted outcomes to estimate the treatment effects. IPW estimates the average treatment effect (ATE) and the potential-outcome means (POMs) from observational data by obtaining probability weights to correct for missing data pertaining to the potential outcomes. Finally, AIPW estimators have the double-robust property as they combine aspects of regression-adjustment and inverse-probability-weighted methods.
The results of treatment effect estimations for revealed son preference show that the corresponding Average Treatment Effects (ATE) obtained for all four birth orders are negative and significant. The ATE of early marriage on fertility stopping at given parity for the four treatment effect estimators range between 9.6 and 13.8 per cent (PSM), 10.5 and 12.4 per cent (RA), 9.2 and 12.3 per cent (IPW), and 9.2 and 12.3 per cent (AIPW), respectively. These findings are highly similar, both in sign and significance, to the baseline estimates, and give strong evidence in favour of higher revealed son preference among early marrying women in Pakistan. Likewise, the results of treatment effects estimated for stated son preference are similar to the baseline estimates.
After the treatment effect estimations, the balancing of the treatment groups is checked using Kernel density plots. The covariates of the two groups are found to be well balanced.

Alternative female marriage indicator
We further check the robustness of our estimates by employing an alternative indicator of female early marriage. We use woman's age at marriage as a count variable rather than a binary indicator. We again come up with significant impact of women's marriage age on the revealed and stated preference for boys. A one-year increase in women's age at marriage is associated with 1-2 per cent higher likelihood of stopping childbearing without a son. Likewise, a oneyear delay in women's marriage in 0.4 (outcome 2)-0.7 per cent (outcome 1) lower stated preference for boys when estimated without controls. The significance of the impact on stated preference disappears, however, when the full set of controls is included.

Alternative definition of female early marriage
Pakistani law sets the female marriageable age at 16 years as against 18 years for men. About 19 per cent of the women interviewed in our survey got married before turning 16. These women are on average less educated and come from poorer households, and can be expected to show greater preference for boys than late-marrying women. The results of the revealed son preference model show that the coefficients of the early marriage variable are highly significant and retain the negative sign. The marginal effects for the first four birth order vary from 7 to 10.2 per cent.
Similarly, the marginal effects for the stated son preference models range between 1.2 and 7.5 per cent, respectively.

Sex-selective abortion
A possible threat to our estimation could be from sex-selective abortion. As discussed in Section 4.1, there is little evidence suggesting widespread practice of sex-selective abortion in Pakistan. Reliable data are scarce, as women are reluctant to report abortions given the social stigma attached to the practice. We explore this possibility through two strategies: First, in our pooled sample, mothers reported the deaths of 16,198 children. Out of these, 7930 (4540 boys and 3390 girls) were reported to have died at zero month. This number may presumably include abortions in addition to still-births and neo-natal deaths. A significant association between death at birth of a female child and mother's early-age marriage may suggest the presence of a discriminatory practice. We fail to find any such association. The coefficients of the estimates are invariably insignificant, with p-values in excess of 0.5.
Second, we limit our sample to the women who do not report any child death, thereby precluding any cases of gender-specific abortion misreported as miscarriage or still birth. We again obtain results similar to our baseline estimations.

Conclusion
The fifth Sustainable Development Goal (SDG) of the United Nations which deals with gender inequality calls for entailing women and girls equal rights to economic resources, and ensuring their full participation at all levels in economic decisions (UN, 2015). A prerequisite to achieving the goal of women's economic empowerment is to eliminate harmful practices, such as marriage before age 18. In this study, we compared the reproductive behaviour and fertility preferences of early-and late-married women using data on married women from four PDHS surveys. The findings of this study underscore the role played by early marriage in altering the gender-specific attitudes prevalent in the society. There is substantial evidence for disproportionate son preference prevalent among early-married women. This evidence points in the direction of gender-inequality traps. Previous research from Pakistan has shown an association between women's say in household decision making and son preference (Javed & Mughal, 2019). Likewise, early-marrying women, who are themselves victims of existing patriarchal customs, help perpetuate gender bias through son-preferring reproductive practices.
These son-preferring norms have non-negligible social and demographic consequences. The desired sex ratio is increasing, sex ratio at last birth is worsening, and son preference's association with modern contraceptive use has become stronger. This can aggravate existing gender gaps in children's anthropometric, health and development outcomes. Tackling these traps requires policy interventions aimed at empowering women. Raising the female marriage age to 18 years could be one option. However, merely passing laws against child marriage is not sufficient to end the practice in the developing countries (Wodon, Tavares, Fiala, Nestour, & Wise, 2017). It is equally important to provide the parents better incentive structures that lead to greater school enrolment and higher labour participation of their daughters. The incidence of female early marriage in Pakistan has decreased over time and the age at first marriage has risen. This demographic transition owes less to any sustained policy initiative or public awareness campaign and more to socioeconomic pressures related to urbanization, improved girls' education and increased female participation in the labour market (Javed & Mughal, 2021). Investing in girls' education can help reduce gender disparities while at the same time delaying marriages, thereby contributing to further reducing the incidence of early-age marriage (Qureshi, 2018).