Where have I got to? Associations of age at marriage with marital household assets in educated and uneducated women in lowland Nepal

Literature review

Findings from previous studies in low-and middle-income countries that have inappropriately used marital household wealth as a proxy for natal household wealth have shown inconsistent findings. In Nepal, for example, Pandey (2017) used marital family wealth and other socio-demographic characteristics (e.g., women’s and husband’s education, caste affiliation, geographical region) to investigate the odds of women marrying before/after 16 and 20 years of age. The results demonstrated no association of wealth with early marriage. In contrast, across South Asia, Scott et al. (2021) found that women from poor rural households were more likely to be married and have children at a young age. However, since wealth was measured in the marital household, these data indicate that poorer households are most likely to recruit in an underage wife, and do not demonstrate that poverty in the natal household drives families to marry their daughters early.

The few studies which correctly measured wealth in the natal household and quantified its association with age at marriage also do not consistently associate poverty with early marriage (Adams & Andrew, 2019; Bajracharya & Amin, 2012; Marphatia et al., 2021c, 2022b; Singh & Revollo, 2016). In our own study population, for example, women’s lower educational attainment rather than natal household poverty was associated with marrying early (Marphatia et al., 2021c).

Other studies that used marital wealth data appropriately have compared the proportion of women married underage by marital household wealth quintiles (Guragain et al., 2017; MacQuarrie, 2016; Male & Wodon, 2016; Mathur, Greene & Malhotra, 2003; Mensch, Singh & Casterline, 2005; Paul, 2019; UNFPA, 2012; UNICEF, 2005, 2021). Analysis of DHS and Multiple Indicator Cluster Survey data (2000–2010) from 78 low-and middle-income countries on women aged 20–24 years found that the poorest 20% of households generally had the largest proportion of wives married underage (UNFPA, 2012). However, as noted by Mensch, Singh & Casterline, 2005, these studies can identify differences in early marriage rates in groups, and monitor changes in prevalence, but do not address why these patterns manifest at the individual level.

In order to understand marriage patterns in more detail, we need to consider the factors that shape the underlying decisions. In the cultural context of the Maithili-speaking Madhesi population of our study, where women have very low levels of autonomy, arranged marriages continue to be the norm (Gram et al., 2018; Morrison et al., 2018). In Nepal in general, some studies report that ‘love’ marriages are increasing. For example, data from 2012 onwards indicate that 32% of people married for love nationwide, and 50% in Chitwan (Allendorf & Thornton, 2015; Human Rights Watch, 2016). However, love marriages remain rare in our study population, and instead, evidence from grey literature suggests that the timing of marriage decisions continues to be shaped by the preferences and characteristics of both natal and prospective marital households (Human Rights Watch, 2016).

In many South Asian societies, marriage decisions relate to several factors (Marphatia et al., 2020a, 2022a; Raj et al., 2014). These include (a) the perceived costs to the natal household of maintaining a daughter, such as providing food, clothes and education; (b) the perceived benefits to the natal household, which include her continuing to undertake unpaid care work; and (c) the perceived benefits to the marital household of gaining a daughter-in-law at a particular age and education level (Human Rights Watch, 2016; Jeffrey & Jeffery, 1994; Maharjan et al., 2012; Mathur, Malhotra & Mehta, 2001; Samuels et al., 2017; Verma, Sinha & Khanna, 2013). Education plays a particularly complex role. A household’s decision to invest in their daughter’s education depends on their desire or economic capacity to send her to school, as even public schools incur material costs, e.g., learning materials, and her reduced contribution to household subsistence. However, women’s education may also be leveraged in marital matches to marry more educated boys (Fafchamps & Shilpi, 2011) and potentially also into wealthier households (Jackson, 2012). In South Asia, though illegal, it is common for the girl’s family to pay a dowry to the husband, and education plays an important role in the bargaining process as the amount payable tends to increase in association with the girl’s age and schooling (Anderson, 2007; Field & Ambrus, 2008; Jeffrey & Jeffery, 1994; Mensch, Singh & Casterline, 2005; Sah, 2012).

In summary, the investigation of how wealth is associated with women’s marriage age is fraught with definitional and data-related issues. In many settings, the girl herself has limited agency over the timing of her marriage. Instead, marriage decisions are largely the consequence of others expressing their own interests. Misinterpreting the source of these interests hinders us from developing a comprehensive perspective on the practice of underage marriage which bridges feminist economics and public health. A feminist economics perspective provides a critical analysis of patriarchal social norms linking women’s low levels of education, unpaid care work and the household economy, all of which shape underage marriage (John, 2021; Urban & Pürckhauer, 2016). Applying this perspective to public health interventions may help to explain why conditional cash transfers have not substantially delayed marriage (John, 2021; Nanda et al., 2015).

Understanding the association of age at marriage with wealth in the marital household is nevertheless an important question, because natal families are very aware of the wealth of a prospective marital household at the time the marriage is being arranged. We have previously used the same study design to investigate how women’s marriage age is associated with the education of the husband (Marphatia et al., 2021a). Again, like wealth, the husband’s education is readily observable to the natal household during marriage negotiations.

Conceptual framework and hypotheses

We cite this previous literature to emphasise that our own study uses marital household assets more appropriately, as a marker of wealth in the household it was actually measured. Our approach allows us to explore a different set of research questions which, compared to previous studies, are more appropriate for data on marital household wealth.

Figure 1 sets our approach in context. Panel A shows that most previous studies have investigated whether poverty in the natal household (inappropriately measured in women’s marital home) is a risk factor for early marriage. Panel B illustrates the goals of our study, which are to investigate whether women’s characteristics at point of marriage (age and education) are associated with the wealth of the households they marry into.

This distinction, between previous studies and our approach, is important and enables us to emphasize that due to methodological flaws, these different questions have in fact been widely confused as detailed below. Conceptually, our study does not investigate where women came from in terms of their circumstances, but rather where they have arrived at marriage. We therefore cannot know if women entered into better circumstances at marriage.

Our study addresses a new hypothesis, which no previous study has systematically investigated:

Hypothesis 1: later marriage age is associated with an increased marital household asset score;

However, as discussed above, Hypothesis 1 requires probing in greater detail, because the association of marriage with marital wealth may not be uniform, but may rather depend on the woman’s level of education. We therefore also test the following hypotheses:

Hypothesis 2: higher educational attainment among women is associated with an increased marital household asset score;

Hypothesis 3: the association of marriage age with the marital household asset score varies according to the level of women’s education;

We first test these hypotheses in the whole sample, and then explore the association of marriage age with assets specifically in uneducated women only, who comprise a substantial proportion of our population. This adds value to our approach because for uneducated women, the association of other predictor variables with marital household asset score is entirely independent of, or unconfounded/mediated by, their education.

If marriage decisions simultaneously consider the age and education of women on the one hand, and the level of assets in the marital household on the other, we might expect that women are married at an age which advantages them economically. We therefore test an additional hypothesis:

Hypothesis 4: for any given level of education, women are most commonly married at an age which would maximize marital household assets.

Variables

Our dependent (outcome) variable was marital household assets, expressed as a continuous score. Each household was assigned an ‘asset score’ based on the ownership of a range and number of consumer goods, and various components of household infrastructure. This standard approach to assets assessment is widely used by nationally representative surveys (Rutstein & Johnson, 2004; Vyas & Kumaranayake, 2006) and has been validated to be appropriate in our study context (MOHP, New ERA & ICF International, 2017). Assets were measured in the household where women were living when they enrolled in the trial, and were therefore measured in either the natal or marital household, but not both. We selected only those households where the woman was living in their marital home, who comprised the majority (78%) of the sample. Since our asset score was measured in the marital household after, and not at, marriage, we controlled for the time since marriage by restricting the sample to women who had married in the past 12 months.

The marital household asset score used in our study was constructed from nine variables using principal component analysis (PCA) (Filmer & Pritchett, 2001; Rutstein & Johnson, 2004; Vyas & Kumaranayake, 2006). Compared to other methods for calculating wealth, the PCA is considered to be the most widely used, robust and comparable approach (Poirier, Grépin & Grignon, 2020; Vyas & Kumaranayake, 2006). These household assets represent relatively stable markers of wealth, relating to durable goods such as the structure of the home or the presence of goods or services that require significant financial outlay. Regarding the marital household, these assets were assumed already to exist before a girl marries into it. We excluded goods such as ownership of a colour television, motorbike, or computer, because they could have been acquired after marriage. The first principal component had positive factor loadings for all nine variables and accounted for 34.5% of the variability, compared to 12.7% from the second principal component. Thus, the first score was taken as the marker of wealth. A review article found that most studies using the PCA wealth index explained 12% to 27% of the variance (Vyas & Kumaranayake, 2006). In comparison, our PCA component 1 explained a higher proportion (34.5%) of the variance, making it a good proxy for household wealth.

The nine variables contributing the highest factor loadings to the first principal component, listed in order of decreasing size (weight shown in parenthesis), were: wall (0.767) and roofing materials (0.747), toilet facilities (0.701), flooring materials (0.700), number of rooms used for sleeping in the house (0.561), land ownership (0.484), access to electricity (0.430), drinking water source (0.399), and use of non-biomass cooking fuel (0.299). These weights reflect the contribution of each variable to the marker of wealth, and imply that the quality of the household’s fabric is the most important aspect in the wealth index. All nine variables had strong positive correlations (p < 0.001) with husband’s levels of education, ranging from 0.184 for roofing materials to 0.402 for toilet facilities. In our study population, assets were mostly male generated because males directly earn and control income. Due to seclusion norms, women are largely confined to the home; being extremely limited in their autonomy, young married women generally do not work for, or control, income (Gram et al., 2017, 2018; Maharjan & Sah, 2012; Morrison et al., 2018). However, women’s unpaid care work (e.g., childcare, household work such as cleaning) nevertheless makes an invaluable contribution to sustaining the household and the well-being of all its members (Marphatia & Moussié, 2013).

Our explanatory (predictor) variables were women’s marriage age and their educational attainment. Our study focused on a highly marginalised group of women, most of whom are illiterate, and therefore cannot precisely tell us their age, nor exactly when they were married. To address this, fieldworkers were trained on determining women’s age as precisely as was possible, by asking numerous questions to verify their response. Women’s marriage age was recorded as an integer value in running years because this was how people generally reported age in this region of Nepal. We converted these ages to completed years (running years minus 1) for analysis. Marriage age was coded into five groups based on the distribution in our data: ≤14 years, 15 years, 16 years, 17 years and ≥18 years. We created these groups because we wanted to know whether later ages at marriage were associated with more assets in the marital home. The youngest and oldest marriage age groups were combined across years due to small numbers. Educational attainment (highest class completed in school) was coded into four levels based on the structure of the education system in Nepal (Ministry of Education Nepal, 2016), and also based on the distribution of data in our sample: none, primary (1–5 years), lower-secondary (6–8 years) and secondary or higher (≥9 years).

Since assets may also be shaped by other markers of social status, our models included husband’s education (coded similarly to women’s education) and caste. In Madhesi society, caste reflects diverse normative behaviours and values, as well as social constraints (Bennett, Dahal & Govindasamy, 2008; Maharjan & Sah, 2012). Caste was measured in the marital home, and, reflecting the distribution of the data, was described as: disadvantaged (Muslim or Dalit), middle (Janjati or other Terai groups), or advantaged (Brahmin or Yadav). We grouped Muslim and Dalit castes together because of their younger marriage age and lower education relative to the middle and advantaged castes. Ideally, we would have included additional variables relating to the wealth of the marital household, such as education of husband’s parents and husband’s (or head of household’s) occupation and income, but our study did not collect these data.

Statistical methods

We used median and interquartile range (IQR) to summarise continuous variables, given the skewed distributions of age-and timing-related data. We tested for biases in characteristics between women whose assets were measured in their natal household versus marital household. We used chi-squared tests for categorical variables. For continuous variables, we used the non-parametric k samples analysis of variance (Kruskal-Wallis test), to test for homogeneity of location. We used scatterplots to show crude associations of the raw marital household asset score against women’s age at marriage, stratified by women’s four educational attainment groups. Spearman’s coefficient described the correlation between women’s age at marriage and marital household asset score by women’s different education groups. To improve interpretation of statistical models, we then standardized the asset score based on the first principal component of the durable assets data to a scale of 0 to 100. Coefficients thus showed the increase in the asset score corresponding to each category of marriage age, education, etc., where the intercept described the mean of women in the reference group for each predictor and the coefficients were deviations in percentage points from that mean, associated with other predictor values or categories.

To test our hypotheses, we fitted four linear mixed-effects regression models with a random effect on the intercept accounting for within-cluster variability. We assessed the models’ assumptions regarding normality and homoscedasticity of the fixed-effects residuals, and of normality of the random effects by inspecting the corresponding plots. Then we tested this formally with the Shapiro-Wilk test, which showed a non-normal distribution of our outcome variable. However, we still presented the linear mixed-effects models in the main article, because quantile regression showed similar results (Tables S3 and S4, Fig. S2). Quantile mixed-effects regression (Geraci, 2014) is an alternative approach. It works on the same scale as the outcome variable, and the interpretation of the coefficients is the same to linear mixed-effects models. The main difference between these two approaches is that the quantile regression estimates effects on the outcome variable’s median (conditional on covariates), whereas linear regression estimates the effects on the mean.

In the full sample, linear mixed-effects models showed the separate associations of marital household asset score with women’s marriage age (Model 1, hypothesis 1) and women’s education level (Model 2, hypothesis 2). For hypothesis 3, we tested the combined interactive associations of women’s marriage age and their education (Model 3), taking into account marital household traits (Model 4). The interactions between maternal age at marriage and their education were defined as the product of the maternal marriage age as a numerical value in five groups, similar to Model 1 (≤14 years, 15 years, 16 years, 17 years and ≥18 years) and the three education groups (‘1–5 years’, ‘6–8 years’ and ‘≥9 years’), relative to uneducated women as the reference group.

Hypothesis 3 also tested the separate associations of marital household asset score with women’s marriage age for only the uneducated women (Model 1), taking into account marital household traits (Model 2). For hypothesis 4, we showed heat maps of the absolute number and percentage of women marrying at different ages by their level of education. This enabled us to examine the most common age for women of a particular education group to marry.

We used linear mixed-effects regression rather than Ordinary Least Squared Regression in order to account for clustering with a random effects term for the 80 geographic (VDC) clusters (which were the unit of randomisation in the trial). Models controlled for women’s age, which, when included together with their age at marriage effectively accounted for the time-gap between when they married and their current age (when they were recruited into the trial). We used women marrying ≤14 years, lack of education and disadvantaged caste as the reference groups. We reported the regression models’ coefficients and their standard errors (s.e.), and statistical significance as: *p < 0.05, **p < 0.01, and ***p < 0.001. We also give the p-value if <0.10. We evaluated goodness-of-fit and report both the marginal and conditional Nakagawa-Schielzeth conditional R². The conditional R² measured the percentage of variance explained by both the model’s fixed and random effects, whereas the marginal R² considered only the variance of the fixed effects (Nakagawa & Schielzeth, 2013). Whilst we reported these R² values for information, we were not trying to explain individual-level variability in household wealth, but rather the overall pattern of association by women’s marriage age, to test whether on average, later women’s age at marriage was associated with greater wealth in the marital household. Models were fitted using the R library lme4 (Bates et al., 2014), lqmm (Geraci, 2014) and SPSS 26 (IBM Corp., Armonk, NY).

Although marriage age, education and assets may vary across trial arms, we did not expect this to bias our results and therefore did not adjust for them in models. This is because the trial recruited already married and currently pregnant women, which means that the interventions could not have influenced marriage age or education. Moreover, assets were measured before the trial was conducted, so the cash supplementation arm could not have changed their value.

Sample selection

Of a total of 24,682 women recruited into the LBWSAT, we first excluded 1,235 women with missing data on assets altogether. Second, of these 23,447 women, we excluded 19,463 women who had been married for ≥1 year. We selected only women within 1 year of marriage to ensure a small time-gap between marriage and when the assets were quantified. Of these 3,984 women, we then excluded 874 women whose assets were measured in their natal home and 8 women who had no information on the household where assets were measured. Our sample for analysis therefore consisted of 3,102 women aged 12–34 years. Table S1 shows differences between women where assets were measured in their natal vs. the marital home, but the magnitudes of effect are small, and are not expected to bias our results.

Description of sample

Our analysis contributes to understanding marriage strategies of the Maithili-speaking Madhesi population in the Nepal Terai. This is important because even among the earlier-marrying and lower-educated women in the Terai, the Maithili-speaking Madhesi women have the highest odds of not being in school, and of marrying before 16 years of age (Marphatia et al., 2020a; Ministry of Education Nepal, UNICEF & UNESCO, 2016; Pandey, 2017). Women in our sample had a median age of 17 years (IQR 2) (Table 1) and the median age at marriage was 16 years (IQR 2). Overall, 20% of women were married ≥18 years, 10% married ≤14 years, and 69% between 15–17 years, while 39% were uneducated, 12% had completed primary, 19% lower-secondary, and 30% secondary or higher education. On average, husbands had completed more education than their wives: 34% of husbands were uneducated, 12% had completed primary school, 19% lower-secondary and 35% secondary or higher education. Within our study population, 31% of marital households were from disadvantaged castes, 44% from the middle, and 24% from relatively advantaged groups.

After standardizing on a scale of 0–100, the median marital household asset score was 39.6 (IQR 32.1). The score was lowest for women who had married ≤14 years (36.9, IQR 25.1), and increased with marriage age, being highest for those who married ≥18 years, (42.2 IQR 38.4). Scatterplots in Figs. 2A–2D show the raw marital asset score against women’s marriage age, stratified by women’s educational attainment, for women who were married between ages 12–24 years to emphasize the association (Fig. S1 shows the entire range of marriage age). There was no linear trend between marital assets and marital age in (a) uneducated women (Spearman’s coefficient 0.01, p = 0.631), those with (b) primary (−0.06, p = 0.272) or (c) lower-secondary education (0.01, p = 0.868), but a positive trend among those with (d) secondary or higher education (0.14, p < 0.001).

Hypotheses 1 to 3

Table 2 presents linear mixed-effects models of the full sample. Focusing only on women’s marriage age (hypothesis 1), Model 1 shows that relative to those married ≤14 years, the asset score increased successively with later age at marriage, most noticeably from 16 years onwards. Compared to women married at ≤14 years, the asset score increased by 1.4% for those married at 15 years, 4.6% for those married at 16 years, 4.9% for those married at 17 years, and 9.1% for those married at ≥18 years. Overall, marriage age (together with women’s current age) explained 10% of the variance in marital household asset score. Focusing only on women’s education (hypothesis 2), Model 2 shows that in comparison to no education, primary, lower-secondary and higher education were associated with 6.8%, 12.7% and 20.3% greater asset score respectively, illustrating a strong dose-response association, as shown in Fig. 3. Overall, women’s education (together with their current age) explained 23% of the variance in marital household asset score.

Model 3 (hypothesis 3) includes both education and marriage age plus interaction terms.

Independent of education, each higher age-group for marriage was associated with 0.6% greater asset score, while compared to no education, completing primary, lower secondary or secondary/higher education was associated with 11.3%, 12.1% and 13.0% greater asset score respectively. However, there were additional interaction effects, whereby taking into account the above associations, each higher age-group for marriage was associated with 1.5% lower asset score for those with primary education, and with 0.2% and 2.1% higher asset score for those with lower secondary or secondary/higher education respectively. This model explained 24% of the variance in asset score.

Figure 4 shows the predicted values derived from Model 3 for marital household asset score by women’s marriage age and education. The model assumes a linear change in assets by marriage age group, and includes random effects by cluster (Table S2). It shows that intercepts of all educated groups were higher than the intercept for the uneducated group. The slope for the uneducated group remained essentially flat, with very small changes across the marriage age groups. The slope for the primary-educated group (1–5 years) decreased until 15 years of marriage, and then remained constant. The slope for the lower-secondary educated group (6–8 years) increased after age at marriage of 17 years, while the slope for the secondary-educated group (≥9 years) showed a constant increase after 15 years. The increase in the asset score was almost 15% higher among women in the highest group of education compared to those who were uneducated for marrying at ≤14 years, and nearly 20% higher for marrying ≥18 years.

Table 2, Model 4 (hypothesis 3) shows the overall pattern remains after adjusting for husband’s education and caste affiliation. The main difference in this model compared to Model 3 is a decrease in the coefficients for the three women’s education groups, indicating that part of the association of women’s education with household assets was mediated by husband’s education. Caste had a minimal effect and was statistically non-significant. Independent of education, each higher age-group for marriage was associated with 0.7% greater asset score, while compared to no education, completing primary, lower secondary or secondary/higher education was associated with 8.6%, 7.2% and 7.6% greater asset score respectively. The interaction effects, taking into account the above associations, indicate that each higher age-group for marriage was associated with 1.5% lower asset score for women with primary education, and with 0.3% and 1.3% higher asset score for those with lower secondary or secondary/higher education respectively. This model explained 29% of the variance in marital household asset score.

Table 3 Model 1 (hypothesis 3) focuses on only the uneducated women. It shows that relative to marrying ≤14 years, marrying at 15, 16, 17 and ≥18 years was associated with 2.7%, 4.9%, 3.2% and 6.8% greater marital asset score respectively. This model explained only 5% of the variance in marital household asset score. Model 2 includes husband’s education and caste. Relative to marrying ≤14 years, marrying at 15, 16, 17 and ≥18 years was associated with 1.5%, 4.4%, 2.4% and 6.2% greater marital asset score respectively. Independently, each higher level of husband’s education and caste affiliation was associated with a greater household asset score. Therefore, the independent effect of women’s own marriage age was partly mediated by marrying more educated husbands and /or those of more advantaged caste. This model explained 12% of the variance in marital asset score.

Overall, Table 2, Model 1 supports our first hypothesis, that women’s later marriage age was associated with greater marital household asset scores. Model 2 supports our second hypothesis, that higher levels of education among women were associated with greater marital household asset scores. Models 3 and 4 support hypothesis 3, that the association of women’s marriage age with the marital household asset score varied according to their level of education. This finding is due to the highly educated group showing a steeper increase in marital assets with marital age, compared to all other groups. Table 3 also supports hypothesis 3, by showing a very modest benefit in terms of asset wealth for delaying marriage for uneducated women beyond 16 years of age. Compared to marrying at 16 years, marrying at the earliest ‘legal age’ of ≥18 years was associated with a small additional payoff of ~1.8% greater asset score.

Hypothesis 4

The assumption that each educational level is associated with a particularly common age at marriage is based on previous research suggesting a direct trade-off between marriage and schooling. For example, if a girl is in school, she is less likely to be married (Sekine & Hodgkin, 2017). Moreover, women who have attained more education tend to marry at an older age (Guragain et al., 2017; Raj et al., 2014; Sekine & Hodgkin, 2017; Wodon et al., 2017). Figure 5 presents heat maps illustrating (a) the overall numbers and (b) distributions of women getting married at a given age by education level. Light blue shaded areas indicate low numbers/percentages, and dark blue shaded areas the highest numbers/percentages.

Heat map (A) shows that within the whole sample, women either completed lower-secondary or more education (≥6 years), or were uneducated, while few had completed only primary education. The common pattern was for women to be uneducated and to marry at 16 years, followed by secondary or higher educated women marrying ≥18 years. There are two potential explanations for the 50 women who had completed ≥9 years of education and were married at ≤14 years. First, they had enrolled in school at a young age, at ≤5 years, as found in other studies (Marphatia et al., 2020b). Second, women who were married at ≤14 years had continued their education after marriage, hence completing ≥9 years of schooling. Looking in more detail within education groups, heat map (B) shows that uneducated women tended to get married at 15–16 years, those with primary education at 16–17 years, those with lower-secondary education at 15–16 years, and those with secondary or higher education at 16 years and ≥18. This threshold effect of secondary education, for marrying ≥18 years, has also been found in other studies (Malhotra, Pande & Grown, 2003; Marphatia et al., 2020a; Pandey, 2017; Raj et al., 2014).

These results provide partial support for hypothesis 4, namely that it was broadly most common to marry around the age when the maximum economic pay-off for the woman had been attained. For the highly educated group, marriage was commonest as expected at ≥18 years, when the economic benefits were greatest. For all other groups, the commonest age at marriage was around 15–16 years. We suggest that this pattern arises because, for all those women who were not highly educated, there would have been a negligible economic pay-off for delaying marriage past this age.