Quantifying the association of natal household wealth with women’s early marriage in Nepal

What do we know about poverty’s association with early marriage?

To date, most studies which investigate the link between wealth and age at marriage have focused on women after they have married. Clearly, it is difficult to measure natal household wealth after girls have already married and moved to their marital households. However, these studies assume that marital household assets are a proxy for the natal household’s assets (wealth) without appropriate supporting evidence. Studies relating to Nepal vary in whether daughters from the poorest (Guragain et al., 2017), or the wealthiest (Aryal, 2007) households have the highest risk of marrying early. Other studies find an inconsistent relationship between wealth and marriage age (Raj et al., 2014) or no relationship at all (Pandey, 2017).

Some studies have used household wealth data appropriately, but in two different ways. One group of studies has described the proportion of women married <18 years stratified by marital wealth quintiles without making an assumption that this also represents wealth in the natal household. These studies consistently find that in impoverished regions of the world, including Nepal, the poorest marital household quintile has the largest proportion of women married <18 years (ICRW, 2006; UNFPA, 2012; UNICEF & UNFPA, 2017; MacQuarrie & Juan, 2019). Fewer studies have first measured wealth in the natal household itself, and then investigated if this is associated with the likelihood of women marrying early (Muchomba, 2021). In India, one study found that girls from poorer natal households were more likely to marry early (Singh & Espinoza Revollo, 2016), whereas another study, which measured wealth at several time-points from the daughter’s birth through to adolescence, found no association between wealth and early marriage (Marphatia et al., 2021b). For Nepal, we could only find one such study, which produced a less consistent pattern, whereby it was not the poorest, but the second poorest households whose daughters were married earliest (Bajracharya & Amin, 2012). Collectively, these studies suggest that the relationship between poverty and early marriage is not as strong when wealth is measured in natal households, and that using wealth in the marital household as a proxy for wealth in the natal household may be inappropriate.

Many of the insights on how poverty may push girls into early marriage come from qualitative literature, where household wealth is neither measured nor quantified. Nevertheless, such work does offer valuable insights on subjective perspectives and decision-making around marriage. Several studies from Nepal have identified a strong economic rationale for natal households to marry daughters at a young age, related to material poverty: specifically in relation to the costs of caring (food, clothes) for daughters and paying for their education (Verma, Sinha & Khanna, 2013; Chaudhuri, 2015; Human Rights Watch, 2016; Samuels et al., 2017). The custom of dowry (illegal, but typically still paid from the natal to marital household) is also likely to contribute to early marriage. Dowry generally demands a substantial proportion of a household’s income, and this is greater for families with several daughters (Sah, 2012; Pesando & Abufhele, 2019). As dowry tends to increase with age and education level, it may lower household investment in daughters schooling and also incentivize early marriage (Sah, 2012; Hodgkinson, 2016; Human Rights Watch, 2016; Karim, Greene & Picard, 2016).

Hence, in the poorest families, girls may be perceived as an economic burden and thus the earlier they are married, the better the natal household’s economic welfare (Hodgkinson, 2016; Guragain et al., 2017). However, one study from Nepal reporting both attitudinal and quantitative data found that the primary drivers of early marriage were family pressure, socio-cultural norms, low education and food insecurity; income poverty was cited as a less important factor (Maharjan et al., 2012).

Other markers of disadvantage and early marriage

Beyond low material assets, other markers of disadvantage may also be associated with early marriage. For example, education is a key factor associated with marriage age, and studies generally find that girls’ lower educational attainment (years of schooling completed) increases their risk of marrying early (Raj et al., 2014; Delprato et al., 2015; Sekine & Hodgkin, 2017; Marphatia et al., 2020; Scott et al., 2021). The association between poverty and early marriage may therefore vary by girls’ education level, and poverty may also contribute to whether girls are educated in the first place, because of the costs associated with schooling (e.g. fees, books, uniforms, etc.) (Verma, Sinha & Khanna, 2013; Chaudhuri, 2015; Samuels et al., 2017).

To understand the association between poverty and the timing of women’s marriage, we also need to consider other relevant socio-economic factors. In rural contexts, agrarian land-holding is another relevant marker of household wealth (Fisher & Naidoo, 2016). Landlessness may increase the risk of food insecurity, which in turn has been associated with both lower schooling and earlier marriage (Moock & Leslie, 1986; UNICEF, 2014). Caste affiliation is also linked to socio-economic status, with girls from disadvantaged castes generally completing less education and also marrying <18 years (Stash & Hannum, 2001; Sah, 2018; Devkota, Eklund & Wagle, 2020). There is much less literature on the role of the natal household’s geographic location in relation to early marriage (Marphatia et al., 2021c), but greater distance to school has been found to be a key constraint to accessing education (Jamison & Lockheed, 1987; Ayral, 2014; Devkota & Upadhyay, 2015). Thus, if schooling is not a viable option, marrying daughters early may alleviate household financial pressures and food insecurity (Maharjan et al., 2012; Human Rights Watch, 2016; Samuels et al., 2017).

Socio-cultural norms are also likely to shape both the timing of marriage and the amount of education that girls are likely to complete. Bicchieri, Jiang & Lindemans (2014) define these normative social preferences as ‘moral rules’ that govern decision-making relating to women’s life options, whether they refer to marriage, chastity, education, employment, etc. Failing to conform to these norms may adversely affect a girl’s marital options, and also her natal household’s social standing in the community (Caldwell, Reddy & Caldwell, 1983; Maertens, 2011, 2013). However, such norms can also change over time. Several studies have suggested that secular changes in attitudes and norms, coupled with widespread advocacy for minimum marriage age legislation, improvements in household wealth, and increased girls’ educational attainment, are collectively likely to explain the overall decrease in early marriage over the past ~15 years across South Asia (Raj, McDougal & Rusch, 2012; Allendorf & Thornton, 2015; MacQuarrie & Juan, 2019; Prakash et al., 2020; Scott et al., 2021).

Study aim and hypotheses

Our study aims to contribute new insights on the economic and social drivers of women’s early marriage and their lack of education in low-income settings. Using objective data on wealth and broader markers of disadvantage measured in the natal household on 2,432 women aged 18–39 years from lowland rural Nepal, we investigate whether natal household poverty is associated with marrying early. We define ‘early marriage’ using several different age groupings, because most women in our population marry well below the 18-year threshold (universal minimum legal age) conventionally used to define early marriage. This is crucial because if we only use the 18-year threshold, we might miss identifying the factors associated with variability in age at marriage as it is experienced in this early-marrying population.

Since two-thirds of our sample are entirely uneducated, we also investigate whether poverty is associated with early marriage in uneducated women, and whether poverty is associated with women being uneducated in the first place. The uneducated women are interesting to examine on their own because for this group, variability in education cannot confound the association between wealth and marriage age. Our models adjust for women’s age, to capture potential cohort effects and which may indicate secular changes in social norms over time. To ensure observed associations between poverty and early marriage are not an artefact of related factors, we include women’s education level as another key exposure, and also broader markers of socio-economic disadvantage measured in women’s natal household: agrarian land-holding, geographic location and caste affiliation.

We investigate four hypotheses:

1. that natal household poverty is associated with marrying early, using different ages to define this outcome: <15, <16, <17 and <18 years, in each case compared to marrying ≥18 years;

2. that women’s lower educational attainment, independent of natal household poverty and broader markers of socio-economic disadvantage, is associated with early marriage: <15, <16, <17 and <18 years, in each case compared to marrying ≥18 years;

3. that amongst uneducated women, poverty, independent of broader markers of socio-economic disadvantage, is associated with early marriage at different ages;

4. that natal household poverty, independent of broader markers of socio-economic disadvantage, is associated with women being uneducated.

Data

Outcome variables

For our first outcome variable, women’s ‘early marriage,’ we use several different age groupings since the majority of our sample (87%) married below the UN stipulated minimum age of 18 years. In Nepal, the legal minimum age at marriage is 20 years, and until recently, marriage at 18 years was possible with parental permission (His Majesty’s Government of Nepal, 1963; Government of Nepal, 2017). Since our trial was conducted prior to this change in legislation, and as few women had married ≥20 years, we use the minimum marriage age cut-off of 18 years as the reference group. To ensure comparability across these results, the same reference group, marrying ≥18 years, is used irrespective of the age used to define early marriage. We examine the factors associated with marrying <15 years (hence excluding women who married between 15–18 years), <16 years (hence excluding women who married between 16–18 years), <17 years (hence excluding women who married between 17–18 years), and <18 years of age (includes the full sample). In other words, each of the models above the <15 years has the lower group nested within it, but excludes the higher group up to 18 years. Figure 1 illustrates our approach.

Our second outcome variable, women ‘being uneducated,’ is coded as any formal education (≥1 year) vs no education (0 years). We use this cut-off because two-thirds of our sample have never been to school.

Statistical methods

We first test for bias in the characteristics of women with assets measured in their natal versus marital households using chi-squared tests (categorical variables) and non-parametric k-sample analysis of variance (Kruskal–Wallis test; continuous variables). We also test for differences in individual assets by asset quintile level using chi-squared tests (categorical variables) and ANOVA (reporting the mean and standard deviation, SD). Given the skewed distribution of women’s age, we report the median (and interquartile range, IQR) values in completed integer years. A heat table examines the distribution of women by education level and natal household wealth quintiles. We use SPSS 26 to conduct these analyses (IBM Corp., Armonk, NY, USA). Using R library tidyverse and ggplot2, we create boxplots to stratify the association of women’s marriage age with their education by natal household asset quintiles (Wickham, 2016; Wickham et al., 2019).

We fit logistic mixed-effects models with a random effect on the intercept accounting for within-cluster variability. If i denotes individuals nested within geographic clusters, the equation for a general regression model of this form with p covariates, and random effects in the intercept ( $u_{i0})$ and in one covariate $(u_{i1})$ is shown below. Logit is the log-odds function, defined as $\mathrm{l}\mathrm{n}\left(\frac{{\pi }_i}{1-{\pi }_i}\right)$, where π_i is the conditional probability that the binary outcome variable Y_i equals one divided by the probability that it equals zero. ${\beta }_0$ is the intercept (constant term) and ${\beta }_1$ is the slope, or coefficient estimate describing the relationship between the outcome and the predictor variable $X_1$. The index j denotes variables, respectively, and the residual error term, ε_i is assumed to be Normally distributed with mean 0 and variance σ ², whilst the random effects are assumed to follow a bivariate Normal distribution with mean (0,0) and variance-covariance matrix $(\begin{array}{c}{\vartheta }_0^2{\tau }_{01}\\ {\tau }_{01}{\vartheta }_1^2\end{array})$, independently of ε.

Our logistic mixed-effects model, shown below differs from the usual fixed-effects model in that it includes random effects accounting for unobserved heterogeneity due to the 80 geographic clusters of the trial:

$$\mathrm{l}\mathrm{o}\mathrm{g}\mathrm{i}\mathrm{t}=\ln \left(odds\right)=\mathrm{l}\mathrm{n}\left({\displaystyle \frac{{\pi }_i}{1-{\pi }_i}}\right)=\left({\beta }_0+u_{i0}\right)+\left({\beta }_1+u_{i1}\right)x_{i1}+\sum _{j=2}^p{\beta }_j{x}_{ij}+{ϵ}_i$$

Our models estimate the probabilities, derived from adjusted Odds Ratios (aORs) with 95% Confidence Interval (CI) of women (a) marrying <15, <16 <17 or <18 years, and (b) being uneducated. Models of the factors associated with being uneducated do not include marriage age because it is not appropriate to use a factor which occurred at time B (marriage age) to predict something earlier at time A (never starting school). All of the models control for women’s age (potential cohort effect), which, when included together with their age at marriage effectively accounts for the time-gap between when they married and their current age (when they were recruited into the trial). Since our interest is in understanding whether poverty (defined relative to the richest quintile) is associated with early marriage, the reference group is set as the highest category across variables, hence: richest asset quintile, secondary education, agrarian land-holding of ≥1 hectare and advantaged caste. Living near to the biggest bazaar (<30 min) is set as the reference group because we assume it is a proxy for better access to school and other resources.

We evaluate goodness-of-fit using the Nakagawa–Schielzeth marginal R² which measures the percentage of variance explained by the model’s fixed effects (Nakagawa & Schielzeth, 2013). Logistic mixed-effects models are fitted using the R library lme4 (Bates et al., 2014).

We adjust for trial arm because women from natal households were more likely to enroll in the cash and food interventions. However, since the trial recruited women who were already married and currently pregnant, the intervention could not have influenced marriage age or education (which typically is ended before/once women marry). Moreover, assets were measured before the trial was conducted, so the cash supplementation arm could not have changed their value. As the trial arm was not associated with our outcomes, we do not report the findings, although it is still controlled for in our analyses.

We conduct a sensitivity analysis to examine if the association between poverty and our outcomes change if we apply looser selection criteria, by including all of the women measured in their natal household (n = 3,379) regardless of their age. While this introduces the possibility of distortion due to selection bias (as discussed above), it nevertheless allows us to test our hypotheses with a much bigger sample of women.

Sample selection

Married pregnant women were interviewed in the home in which they were residing at the time of recruitment into the trial. Of the 25,090 women recruited into our study, we first exclude 408 women with multiple pregnancies during the trial to ensure they are not double counted in our analysis. Second, of the remaining 24,682 women (raw data in Data S1 contains this sample), we exclude 3,968 women who had no data on the household in which assets were measured and a further 17,335 women who were interviewed in their marital home. This leaves 3,379 women whose assets and other characteristics were measured in their natal household.

Third, we exclude 947 women aged <18 years because they would not have had the chance of marrying at older ages, nor adequate time to finish greater levels of education before marrying. The relationship between wealth, education and marriage age is not likely to be distorted due to the selection bias of recruiting only married pregnant women into our study. In the context of our study, women are very unlikely to continue their education after marriage (Sekine & Hodgkin, 2017). The vast majority of married women are also likely to have children, and it is common in Nepal get pregnant in the first 2 years after marriage (MacQuarrie, 2016; Marphatia et al., 2020). Our analysis therefore includes 2,432 women aged 18-39 years measured in their natal households, representing 9.7% of the total 25,090 women recruited into the trial.

Table S1 examines whether the characteristics of the women measured in their natal homes are different to those measured in their marital homes. This enables us to identify potential bias in our sample which may result in these two groups of women having a different relationship between wealth and marriage age. A direct comparison between these women is not possible because we do not know the socio-economic background of the natal households of the women who were measured in their marital home.

In comparison to women who were measured in their marital home, those measured in their natal homes are younger, have been married for less time, but at an older age. A greater proportion of women measured in their natal households are from disadvantaged castes, residing further away from big bazaars, and in the trial’s cash and food supplementation arms. It is possible that the trial may have incentivised women to return to their natal homes to access these incentives. Women did not differ in their education level, age at first pregnancy, or household asset score. Whilst these results show that women who went to their natal homes during pregnancy are different than those who stayed in their marital homes, the differences are small. The women in our sample still represent an important sub-group of women with rare data on the socio-economic characteristics of their natal homes.

Description of sample

Table 1 describes our sample of women aged 18–39 years. Women are of median age 21 years (IQR 4) and have been married for a median of 5 years (IQR 5). Marriage is typically early among women (median 15 years, IQR 3), and 18% married ≥18 years. Two-thirds of the women are uneducated, 39% are from natal households without agrarian land and disadvantaged castes respectively, and 29% live far from large bazaars.

The heat table shows the overall number of women by their education level and natal household wealth quintiles (Fig. 2). Green shaded areas indicate low numbers and red shaded numbers the highest numbers. Within the whole sample, uneducated women are most likely to come from poorer natal households, and the more educated women are more likely to come from richer households. However, since two-thirds of our sample is uneducated, there is substantial variability by wealth that is independent of education.

Figure 3 illustrates the association of women’s marriage age with their education level, stratified by natal household asset quintiles. Overall, within each wealth group, women with more education marry later. The level of education associated with delayed marriage differs slightly across wealth groups. Interestingly, even in the poorest and poor wealth groups, the median age at marriage is 2 years later among women with ≥6 years of education than among those with no or less education.

Hypothesis 1

Hypothesis 1 investigates the association of the natal household asset score with the odds of early marriage, using each of the four age-definitions of early marriage. Across Models 1–4 in Table 2, there is no evidence of a cohort effect, which may be explained by the relatively narrow age range of our sample. Relative to the richest asset quintile, women from all other asset quintiles have an elevated risk of early marriage, regardless of the age threshold used to define ‘early’. However, there is little consistent gradient between the coefficients of the poorest four asset groups, indicating that the main difference in marriage age is between the richest and the rest (this is confirmed by analyses using other asset groups as reference (not shown)). The variance in marriage age explained by the four models is very low: 5.0%, 3.2%, 2.8% and 2.1%. These results do not support our first hypothesis, that natal household poverty is associated with the likelihood of marrying at different early ages.

Hypothesis 2

Hypothesis 2 investigates the association of broader socio-economic factors with the likelihood of marrying early as defined by the four different age thresholds. With the exception of Models 3 and 4, age is not a significant factor associated with early marriage, indicating the absence of a cohort effect (Table 3). Models 3 and 4 shows that older women are less likely to marry <18 years than ≥18 years, which is perhaps surprising: we expected that further in the past (older women) prevailing socio-cultural norms might have more strongly favoured earlier marriage. This cohort effect only emerges when education is controlled for in the analysis, which suggests that it is connected to difference in the availability of, and attitudes towards, education over time. It is also possible that this cohort effect emerges because in the past, high levels of education were rarer, and therefore more tightly associated with delayed marriage compared to more recent times, when education is more widely available to the younger women in our sample.

Across Models 1–4 in Table 3, the asset score is no longer associated with early marriage when women’s education is included. The effect of lower wealth appears therefore to be channelled through women’s education, which is a better predictor of early marriage in this population. There is a clear education gradient in the likelihood of marrying early, however ‘early’ is defined. Relative to secondary education, all other education levels have an elevated risk of early marriage, with uneducated women having a substantially greater risk. The magnitude of effect of education also decreases with each additional year used to define early marriage, suggesting that education tends to increase with age, but assets do not. Neither land, caste nor geographic location are associated with any of the marriage age models when women’s education is controlled.

Compared to Table 2, these models are almost three times better at explaining the variance across the four marriage age groups: 23.5%, 16.4%, 13.1% and 10.6%. These results support our second hypothesis, that women’s lower educational attainment, independent of natal household poverty and broader markers of socio-economic disadvantage, is associated with early marriage, whatever the age threshold used to define ‘early’.

Hypothesis 3

Hypothesis 3 investigates whether amongst the uneducated women (n = 1,610), poverty, independent of broader markers of socio-economic disadvantage, is associated with early marriage at different ages. Results show no cohort effect, nor any association with wealth (Table 4 Models 1–4), suggesting that the uneducated women who marry early are not systematically more or less wealthy compared to those marrying ≥18 years. There is also no association between broader socio-economic factors and early marriage, whatever age threshold is used to define early marriage (Table 5 Models 1–4).

In Table 4, models explain 6.2%, 5.4%, 5.0% and 5.0% of the variance in women marrying <15, <16, <17 years and <18 years respectively. In Table 5, models explain slightly more of the variance in early marriage across the age groups 7.1%, 6.6%, 5.8% and 6.0%. In comparison to models in Table 3 for women with all educational levels, those in Table 5 explain less variance. The total absence of women’s education in these models is likely to explain this difference. These results do not support our third hypothesis, that among uneducated women, poverty, independent of broader markers of socio-economic disadvantage, is associated with early marriage at different ages.

Hypothesis 4

Hypothesis 4 investigates whether poverty, independent of broader socio-economic factors, is associated with women being uneducated. Across the two models in Table 6, older women are more likely to be uneducated, reflecting the increase in the availability and acceptability of education for girls over the 20-year period during which our sample was maturing.

Model 1 shows a clear wealth gradient in the likelihood of being uneducated. Relative to the richest quintile, all other asset quintiles have an elevated risk of being uneducated, with the poorest quintile having a substantially greater risk. Model 2 finds a similar pattern after inclusion of broader socio-economic factors. In comparison to Model 1, the magnitude of effect of the first wealth quintile is weaker, that of the mid quintile only slightly weaker, and unchanged in the two richest quintiles. Relative to higher agrarian land-holding, both none and some land-holding are associated with being uneducated. Relative to the advantaged caste, the two disadvantaged castes are associated with being uneducated, with the risk substantially greater for the Muslim caste. Relative to living near a big bazaar, living 30–59 min away, but not further distances, is marginally associated with being uneducated.

Model 1 explains 21.0% of the variance in women’s education, which is lower than Model 2, which explains 36.3% of the variance. These results support our fourth hypothesis, that poverty, independent of broader socio-economic factors, is associated with women being uneducated.

Supplementary analysis

A potential reason why poverty may not be associated with women’s early marriage, irrespective of the age used to define early marriage, could be that the individual assets owned by households do not actually differ between the wealth quintiles. Table S2 shows, however, that there is substantial variability across the eight individual assets used to produce our composite asset score using PCA. We also include land-holding in this analysis because it is another marker of wealth in our primarily agrarian population. These individual assets and land ownership matter for daily life and indicate the household’s purchasing power. Our results show that asset ownership does indeed differ by wealth levels, nevertheless our analyses described above show that wealth in itself is not associated with women’s early marriage in this population.

Given the selection bias in our study of recruiting young, already married and pregnant women, we have restricted the analyses described above to women aged ≥18 years only. With the exception of the association of age, our results are nonetheless similar if we include the full sample of women aged 12–39 years (n = 3,379) (Tables S3–S7). The significance of the age variable in these models is a product of the selection effects (as discussed above), and not a cohort effect.

Implications

Our results have implications for research, policy and practice. First, there is an urgent need for informative objective data measuring wealth in women’s natal households, ideally before they marry, and also in women’s marital households, ideally at the time of their marriage. This approach would provide much needed evidence on whether natal household poverty is indeed associated with the likelihood of early marriage and whether women marry into households of similar wealth levels.

These data-related issues, and in particular the inadequate understanding of the association of wealth and marriage age, may partly explain the inconsistent results of interventions targeting poverty as the main driver of early marriage. A recent systematic review found conditional cash transfers (CCTs) to keep girls in school were more effective in delaying marriage than those directly targeting delayed marriage or poverty (Malhotra & Elnakib, 2021). In Bangladesh, CCTs supported girls to stay in school for longer, but delayed marriage for only the youngest girls (aged 12–14 years) living in the poorest district (Amin, 2007). In India, another intervention gave CCTs to parents if their daughters had reached the age of 18 years unmarried. Compared to a control group, the girls who participated in this intervention stayed in school up to grade eight, and were in fact more likely to marry just after the 18 year cut-off stipulated by the intervention (Nanda et al., 2015). In both interventions, the cash transfers was primarily used by parents to pay the higher cost of dowries demanded by marital households for older, more educated girls (Amin, 2007; Nanda et al., 2015).

Second, we need to better understand the different factors increasing the risk of marrying at different ages, for women of different education levels. Reducing the costs of schooling (e.g. fees, learning materials, offering scholarships) and improving the quality of education may help girls to stay in school for longer and also delay marriage, thereby achieving several of the Sustainable Development Goals (Muchomba, 2021). Further research is needed to understand whether the expansion of free state provision of secondary education would have a sustained impact on delaying marriage. However, any school-based efforts will miss the girls who never went to school in the first place or have already dropped out and married.

Third, there needs to be a collective shift in societal gendered norms and the value attributed to girls and women in society (Maertens, 2013; Bicchieri, Jiang & Lindemans, 2014; Marphatia, Amable & Reid, 2017). Changing norms is difficult and slow, as shown by social interventions that did not succeed in delaying marriage age in India (Prakash et al., 2019; Ramanaik et al., 2020). Moreover, in populations where women marry very early, it may initially be more realistic to delay marriage by 1 year at a time. In our study, this would mean first delaying marriage from 15 to 16 years, and then from 16 to 17 years. ‘Nudging’ populations towards a slightly later marriage age for women may therefore lead to more substantial secular changes over time. Whilst this may sit uncomfortably with the human rights constituencies advocating for the 18-year minimum marriage age, ignoring these practiced norms may render us even further from the common goal of delaying marriage overall (Schaffnit, Urassa & Lawson, 2019). Delaying marriage, even by 1 year, will inevitably delay the age at first child bearing. Early marriage must therefore be seen as a critical concern for public health (Marphatia, Amable & Reid, 2017).

Limitations

Our study has some limitations. With cross-sectional data, we can only investigate associations and not causality. Although we include education in early marriage models, like other studies, we do not know the direction of this association. We did not have data on the amount of dowry paid by women’s natal household, nor whether it accelerated the timing of their marriage. Our study involved only married pregnant women and we found that women measured in their natal homes during pregnancy differed from those who remain in their marital homes. However, these differences were small. As we do not have matching data on pregnant women in their marital homes we cannot explore the other ways in which they may differ. In our population, women are also very unlikely to continue their education after marriage, and the vast majority of married women are likely to have children, and it is common in Nepal get pregnant in the first 2 years after marriage. Our asset score was measured in the natal household after, and not at, marriage. To address this, our asset score excluded items which could have been purchased after marriage. Despite these limitations, our study benefited from a large sample size and the unique data on several different socio-economic variables as well as women’s education. The associations between poverty, education and marriage age, and poverty and lack of education identified in our study are likely to be widely applicable, especially to similar Madhesi populations living around the bordering regions of India and Nepal.