1 INTRODUCTION

Preterm delivery refers to a delivery occurring before 37 weeks gestation (Althabe et al., 2012). Of the 15 million infants delivered preterm globally each year, an estimated 1 million die before the age of five years, making it the largest cause of child mortality worldwide (Chawanpaiboon et al., 2018; Lawn et al., 2005). Reducing preterm delivery rates is therefore a vital component of reducing infant and child mortality and morbidity, contributing towards the United Nations Sustainable Development Goal (SDG) 3, which aims to end all preventable deaths under the age of five years by 2030 (United Nations, 2020). From a physiological perspective, preterm delivery interrupts the continuous supply of nutrients via the placenta during a critical period of development (Czech-Kowalska, 2020). Meeting nutritional demands through enteral feeding after birth is constrained by the immaturity of the gut in preterm infants, meaning they are prone to poor growth and development in early life (Zozaya et al., 2019). Furthermore, preterm infants have an increased risk of being exposed to inflammation, either prenatally or postnatally, which impairs neurodevelopment (Cappelletti et al., 2016; Lee et al., 2021). Prematurity therefore has long-term consequences on the health and cognitive development of individuals into adulthood (Kajantie et al., 2010; MacKay et al., 2010; Markopoulou et al., 2019). This can cause families significant psychological and financial hardship (Moster et al., 2008; Rogers & Velten, 2011; Swamy et al., 2008).

Populations in Africa and Asia in particular bear the burden of preterm delivery, as these continents account for around 80% of cases globally (Chawanpaiboon et al., 2018). In South Asia, rates of children dying from complications following preterm delivery are declining at a slower rate compared to other causes of death regionally, meaning that the relative contribution of preterm delivery to childhood mortality has increased over the past 20 years (Liu et al., 2016). Recent efforts have improved the survival of premature infants, however in the Global South, particularly in remote and rural regions, limited access to medical facilities or skilled birth attendance means these improvements are not universal (Choulagai et al., 2017; Iams et al., 2008). Therefore, understanding the risk factors for and mechanisms involved in preterm delivery in these regions are key for prevention and related care.

There are several biological and social factors which have been associated with preterm delivery, such as a young age at pregnancy (Althabe et al., 2015; Ganchimeg et al., 2014; Gurung et al., 2020; Kumar et al., 2018; Stewart et al., 2007). This is very relevant in South Asia where an estimated 30% of girls give birth before their 18th birthday (Scott et al., 2020). In South Asia, marriage is the main context for sexual intercourse and therefore a young age at marriage is a key determinant of the age of first childbirth (Ministry of Health and Population (MOHP), 2017). Child marriage describes a marriage in which one or both spouses are below the age of 18 years, and is a fundamental violation of human rights (UN General Assembly, 1948; United Nations Population Fund, 2012). While the practice has decreased globally over recent decades, it remains prevalent in South Asia, where over half of marriages still take place during childhood, however the prevalence of child marriage varies greatly between regions (Scott et al., 2020; UNICEF, 2018, 2017). Given the scale of the practice of child marriage, understanding the associated consequences is a priority.

Marriage is associated with a range of lifestyle changes and new responsibilities, which present emotional, social and financial challenges, particularly during adolescence (Marphatia et al., 2017; Nour, 2009). The emotional consequences of child marriage are especially severe for girls (Raj et al., 2018). In South Asia, marriage generally requires girls to move in with their new husband's family and take on a new position of deference in the household (Harris-Fry et al., 2018; Marphatia et al., 2017). Marriage also typically marks an end to a girl's formal education, limiting their personal freedom (Field & Ambrus, 2008; Marphatia et al., 2020; Sekine & Hodgkin, 2017). In this setting, girls with a lower level of education have reduced access to antenatal services and are less likely to participate in household decision making, including decisions regarding their own health (Ministry of Health and Population (MOHP), 2012; Sekine & Carter, 2019). Furthermore, those marrying during childhood have an increased risk of experiencing intimate partner violence (IPV; Kidman, 2017). Young brides in South Asia also report lower use of contraception and higher rates of pregnancy termination than those marrying in adulthood, highlighting the adverse reproductive health sequelae (Godha et al., 2013; Raj & Boehmere, 2013). These consequences are not short lived; emotional distress, lack of schooling, reduced access to healthcare, IPV, and poor reproductive health have all been associated with preterm delivery, meaning children born to mothers married during childhood are likely to be more vulnerable to the associated adverse health effects (Baer et al., 2019; Dunkel Schetter, 2011; Efevbera et al., 2017; Franck et al., 2020; Walani, 2020). Despite this, there is a dearth of empirical evidence on maternal health outcomes following child marriage (Godha et al., 2013).

Previous studies have focused on the consequences of child marriage within the context of early childbearing, viewing child marriage as a gateway to early pregnancy (Godha et al., 2013; Mathur et al., 2003; Nasrullah et al., 2014; Paul, 2020; Rahman et al., 2018). While adverse consequences of early pregnancy have been identified, there is considerable debate around whether these are the result of the mothers' biological immaturity, or stresses associated with socio-economic factors or dynamics of the marital household (Gurung et al., 2020; Jiang et al., 2018; Maharjan et al., 2019; Shrestha et al., 2010; Stewart et al., 2007). Most likely, a combination of the two issues is relevant, particularly in cases of child marriage. In South Asia, pregnancy is almost always preceded by marriage and therefore these two exposures are closely intertwined. However, they remain separate events with individual drivers and consequences (Adhikari & Bott, 2003; Marphatia, Wells, et al., 2021). The interval between a girl's age at marriage and first pregnancy differs according to their age at marriage, as well as other factors which may delay the consummation of the marriage or influence their use of contraceptives, such as where they live, their ethnicity, educational status or religion (Gubhaju, 2009; Staveteig et al., 2018). Furthermore, as discussed above, a girl's age at marriage is associated with their access to healthcare, level of education and social engagement, which may impact their reproductive and maternal health through mechanisms independent of their age at pregnancy. Elucidating the consequences of both early marriage and early pregnancy is key to determine where best to focus efforts to improve maternal and child health. Furthermore, previous studies in this field have been limited by small sample sizes and low rates of child marriage (Huang et al., 2021; Pandya & Bhanderi, 2015; Rahman et al., 2018). This study uses data from the Low Birth Weight South Asia Trial (LBWSAT) to investigate the independent associations of age at marriage and first pregnancy with preterm delivery in rural lowland Nepal. Almost 90% of the LBWSAT participants were married during childhood (<18 years), making this dataset particularly well suited to studying its consequences (Marphatia, Saville, Manandhar, Amable, et al., 2021; Marphatia, Saville, Manandhar, Cortina-Borja, et al., 2021). Analysis of primigravida and multigravida women separately examines whether any observed associations are restricted to the first pregnancy, or if longer term sequelae persist in subsequent pregnancies. This analysis also provides an insight into the mechanisms involved in preterm delivery.

2 METHODS 2.1 Study context

This analysis used data from the LBWSAT, a non-blinded, cluster randomized controlled trial conducted in the Dhanusha and Mahottari districts of Nepal. The LBWSAT assessed the impact of community-based participatory learning and action (PLA) women's groups, with and without food or cash transfers, on birth weight and infant weight-for-age z (WAZ)-scores. All married women and girls aged 10–49 years across 80 village clusters were invited to take part in the trial, permitted they or their husband had not undergone surgical family planning (Saville et al., 2018). About 63 308 participants consented to menstrual monitoring from 80 randomized village development committee (VDC) clusters, and 25 090 pregnancies were detected between December 2013 and February 2015 (Saville et al., 2018). VDC clusters were randomly assigned to one of four interventions: behavioral change PLA groups, PLA groups + cash given to pregnant participants, PLA groups + “Super Cereal” supplement given to pregnant participants, or existing government programmes. The detailed study protocol and primary outcome results have been published previously (Saville et al., 2016, 2018; Style et al., 2017).

Informed written consent was taken from all participants in this study, and their guardians if they were aged <18 years. Research ethics approval was obtained from the Nepal Health Research Council (NHRC) (108/2012) and the University College London (UCL) Ethical Review Committee (4198/001) for primary data collection, and from the NHRC (292/2018) and the UCL Ethical Review Committee (0326/015) for the secondary analyses presented in this article.

2.2 Variable selection

The outcome variable for this analysis was the occurrence of preterm delivery (<37 weeks/259 days gestation). Gestation length (GL) was calculated as the time between last menstrual period (LMP) and date of delivery. LMP was determined using maternal recall for 18 306 pregnancies and using ultrasound for 1321 pregnancies where available. Agreement analysis between GL determined by ultrasound and maternal recall was undertaken by calculating Pearson's correlation coefficient and Cronbach's alpha for GL categories. The binary preterm delivery variable was selected rather than a continuous GL variable to align with other research and to minimize the influence of inaccuracies in maternal recall of LMP.

The primary exposure variables for this analysis were the participant's age at marriage and age at first pregnancy. In this trial's setting, most people count their age in running years rather than completed years. Therefore, age at marriage and first pregnancy were collected as integer values in running years and converted to completed years (running years −1) for analysis. To investigate the importance of policy aiming to reduce childhood marriage and motherhood, and based on the distribution and pattern of data in our sample, age at marriage was coded into four groups: ≤14 years, 15 years, 16–17 years, and ≥18 years, while age at first pregnancy was coded into three groups: ≤15 years, 16–17 years, and ≥18 years.

Using a priori knowledge, a directed acyclic graph (DAG) was constructed using DAGitty to identify the minimal sufficient adjustment variables to account for confounding (Textor et al., 2011; Supplemental Figures 1–4). The arrows in the DAG were entered to represent the hypothesized direct causal effect of one variable on another. Variables that are hypothesized to be directly antecedent to the exposure (age at marriage or age at pregnancy) and outcome (preterm birth) are indicated in pink and identified as a confounder (Hernán et al., 2002). This approach identified maternal education, maternal caste group and material household assets as confounders for the association between age at marriage and preterm delivery, as these factors are hypothesized to have a causal effect on both exposure and outcome (Supplementary Figures 1 and 2). For the association between age at first pregnancy and preterm delivery, the same factors plus age at marriage were identified as confounders (Supplementary Figures 3 and 4).

Maternal education was coded into four levels based on the Nepali education system and the distribution of data in our sample: no formal education, primary (1–5 years), lower secondary (6–8 years) and secondary or higher (≥9 years). The household asset score was determined using principal component analysis, reflecting ownership of consumer goods such as a color television, motorbike, or computer, land ownership and household infrastructure (Saville et al., 2018). Maternal caste was grouped into three groups: disadvantaged comprising of Dalit and Muslim, middle comprising of Janajati and other Terai castes, and advantaged comprising of Yadav and Brahmin.

Maternal weight and mid-upper arm circumference (MUAC) were measured in early pregnancy (8–30 weeks). Maternal height was measured in early pregnancy, and endpoint where earlier height was missing. Body mass index (BMI) was subsequently calculated using early pregnancy weight and height collected at either timepoint, and adjusted for gestational age. However, these measurements were only available for a sub-set of women. Season of birth and sex of the baby was determined at birth (within 42 days). Maternal age at current pregnancy was also collected at baseline for additional analysis in multigravida women, categorized into six categories: <18 years, 18–20 years, 21–23 years, 24–26 years, 27–29 years, and ≥30 years.

2.3 Statistical methods

The characteristics of primigravida and multigravida participants were described and compared using chi-squared tests. The characteristics of participants with missing and complete GL data were also compared using chi-squared tests. The relationship between age at first pregnancy and age at marriage was explored visually by generating heat tables in Microsoft Excel (Microsoft Corporation, 2018) and a correspondence analysis plot (“mcaplot” command in Stata). Correspondence analysis represents contingency tables as a low-dimension geometric map of the association, representing the closeness between rows (age at first pregnancy) and columns (age at marriage; Greenacre, 1992).

Mixed-effects logistic regression models were fitted to assess associations of the exposures with preterm delivery, with the VDC cluster included as a random effect term on the intercept (“xtlogit” function in Stata). Minimally adjusted models adjusted for study design, including the random effect for cluster and fixed effects for study arm and strata. Fully adjusted models adjusted for study design plus additional confounders identified using the DAG. Missingness patterns of variables were identified and multivariate imputation using chained equations (MICE) was applied to deal with missing data (“mi” function in Stata; 30 completed datasets; van Buuren & Groothuis-Oudshoorn, 2011). To account for the multi-level nature of the data, the imputation was executed stratifying for each study arm. Models were also run on un-imputed data as a sensitivity analysis. All models analyzed the association between exposures and preterm delivery for primigravida participants and multigravida participants separately.

To determine the potential for mediation by age at pregnancy in the association between age at marriage and preterm delivery, additional models were constructed. These included age at current pregnancy as a covariate in the model for the association between age at marriage and preterm delivery, and removed age at marriage as a covariate in the model for the association between age at first pregnancy and preterm delivery. Further models were also constructed on the subsample of women with data on maternal height, BMI and MUAC to investigate the role of markers of maternal nutritional status. As rates of child marriage are gradually declining over time, additional models including only participants marrying within the last 5 years were constructed to rule out the potential bias of mothers marrying young being older.

All models report adjusted odds ratios (aOR) and 95% confidence intervals (CI). Statistical analyses were performed in Stata IC 16.1 (StataCorp, 2019).

3 DATA CLEANING

Of the 63 308 participants recruited for menstrual monitoring, there were 25 090 pregnancies during the trial period (Saville et al., 2018). Figure 1 demonstrates the cleaning of GL data for this analysis. About 5 394 participants (21%) with missing GL data were excluded. Participants with a GL out by ± ≥365 days (n = 60) were flagged as a data entry error and corrected accordingly, with four excluded due to erroneous dates.

Participant flow diagram showing exclusions undertaken during data cleaning. Data cleaning protocol for this analysis. Briefly, exclusions were made due to erroneous dates, babies having gestational age-adjusted weight-for-age z-scores outside of the World Health Organization z-score cut-offs, erroneous gestation length (<27 weeks/>44 weeks), and multiple pregnancies. Abbreviations: LBWSAT, low birth weight South Asia Trial; GL, gestation length; WAZ; weight-for-age z-scores

Standardized WAZ-scores were calculated using the “zanthro” Stata package which adjusts for gestational age (Vidmar et al., 2013). Analysis of z-score distribution according to gestational age identified a bimodal birthweight distribution for preterm infants, as reported in previous studies (Haglund, 2007; Parker & Schoendorf, 2002). Erroneous GLs were identified as outliers using the World Health Organization (WHO) recommended z-score cut-off of +5 and their GL recoded to term (280 days). Weight-for-age z-scores were then re-calculated and accepted if they were now within range (−6/+5). All participants with a z-score outside of this range were excluded. Additionally, preterm live births with a WAZ-score above +3 who were enrolled >120 days after LMP were excluded due to suspected errors in maternal recall of LMP. Eighty-five pregnancies were excluded due to erroneous WAZ-scores (<1%).

All live births with a GL outside of the 27–44 week range were also excluded as GL was implausible (1456 participants: 6%). Forty-five participants were excluded for having missing data on gravidity. One hundred and thirty-two twin/triplet pregnancies were also excluded. Following data cleaning, 17 974 participants were included in the analysis. Supplemental Table 1 compares participants with complete GL data to those with missing and excluded GL data. Significant differences (chi-squared tests; p < .05) between those with and without missing data were observed according to age at pregnancy, age at marriage, caste group, education level, study arm, and household asset score. Significant differences (chi-squared tests; p < .05) in caste group, education level, study arm, season of birth, sex of the infant, and household asset score were observed between those with and without excluded data.

4 RESULTS

We included 6243 primigravida and 11 731 multigravida participants in this analysis. Table 1 displays and compares the characteristics of primigravida and multigravida participants separately. Twenty percent of primigravida and multigravida participants delivered preterm. 49% of primigravida participants and 63% of multigravida participants had their first pregnancy at <18 years (p < .001), while 86% of primigravida participants and 92% of multigravida participants had married at <18 years (p < .001).

TABLE 1. Study characteristics for primigravida and multigravida participants Primigravida participants included in analysis (%) Multigravida participants included in analysis (%) p-value for difference n = 6243 n = 11 731 Gestational age at delivery Term delivery 80.0 80.4 Preterm delivery 20.0 19.6 .52 Missing 0.0 0.0 Maternal age at first pregnancy ≤15 y 8.4 25.6 16–17 y 40.8 37.7 ≥18 y 50.8 36.7 <.001 Missing 0.0 3.1 Maternal age at marriage ≤14 y 20.3 41.5 15 y 24.2 27.1 16–17 y 41.2 23.3 ≥18 y 14.3 8.0 <.001 Missing 13.8 20.8 Caste three groups Dalit/Muslim – Disadvantaged 32.0 36.3 Janajati/Other Terai castes – Middle 44.3 42.4 Yadav/Brahmin – Advantaged 23.7 21.3 <.001 Missing 0.0 0.0 Household asset score 1 – Most deprived 15.8 22.1 2 18.6 21.2 3 20.4 20.2 4 22.1 19.0 5 – Least deprived 23.1 17.6 <.001 Missing 1.4 1.0 Mother education level Never went to school 49.3 72.9 Primary 11.7 9.5 Lower secondary 15.3 7.1 Secondary and above 23.6 10.6 <.001 Missing 0.2 0.0 Study arm woman enrolled in Control 23.1 22.2 PLA group 23.3 23.7 Cash 27.8 27.6 Food 25.8 26.5 .43 Missing 0.0 0.0 Season of birth Winter: Mid-December to mid-March 31.8 27.7 Spring: Mid-March to mid-June 20.3 22.7 Monsoon: Mid-June to mid-September 25.5 27.4 Autumn: Mid-September to mid-November 22.4% 22.2 <.001 Missing 1.8 2.2 Sex of infant Boy 52.0 53.7 Girl 48.0 46.3 .037 Missing 1.1 0.9 Note: Characteristics of primigravida and multigravida participants, compared using Pearson's chi-squared tests. Abbreviations: PLA, participatory learning and action groups; y, years of age.

Figure 2 illustrates the association between age at marriage and age at first pregnancy in our study population. In the heat tables, green shading indicates the lowest level of association, and red shading indicates the highest. 54% of multigravida participants marrying at ages ≤14 years had their first pregnancy by the age of 15 years, 31% at ages 16–17 years, and 15% aged 18 years or older. 29% of primigravida participants marrying at ages ≤14 years had their first pregnancy by the age of 15 years, 31% at ages 16–17 years, and 40% aged 18 years or older. The correspondence analysis bio plots represent the closeness between age at marriage and age at first pregnancy. The axes demonstrate the proportion of variance explained in a principal components analysis; for primigravida participants, the first and second axes concentrate 59% and 41% of variability, for multigravida they concentrate 76% and 24%. As both variables are ranked by age along the x axes, the first correspondence component reflects increasing age. The second correspondence component has no clear interpretation, reflecting variability in the timing between marriage and first pregnancy.

Heat tables for the association between age at marriage and age at first pregnancy. Heat tables for the association between age at marriage and age at first pregnancy for primigravida participants and multigravida participants in (A) absolute numbers, and (B) as a percentage within age at marriage. Correspondence analysis is represented in the bio plots (C). The % on the x and y axis represent the % of variance explained in a principal components analysis. Abbreviation: y, years of age

Table 2 presents preterm birth rates per 1000 deliveries by categories of age at marriage and age at first pregnancy, showing higher unadjusted rates of preterm delivery among primigravida participants marrying young compared to multigravida. Figure 3 displays the aOR for the association between age at marriage and preterm delivery. Primigravida participants married at ages ≤14 years had significantly higher odds of delivering preterm than those married aged ≥18 years in the minimally adjusted models (Figure 2; aOR 1.45, 95% CI: 1.15–1.83, p = .002). This significance was maintained when adjusting for confounders (aOR 1.28, 95% CI: 1.01–1.62, p = .041). Inclusion of age at pregnancy as a potential mediator slightly reduced the significance of the association between marrying ≤14 years of age and preterm delivery, but did not change the effect size (aOR 1.29, 95% CI: 1.00–1.68, p = .054). While there were trends towards a higher odds of preterm delivery for primigravida participants married aged 15 or 16–17 years, these associations were insignificant for all models. There were no significant associations observed between age at marriage and preterm delivery when analyzing multigravida participants. Results were not changed by the use of multivariate imputation, as shown by the estimates from nonimputed models in Supplemental Table 2.

TABLE 2. Rates of preterm delivery according to each category of age at marriage and age at first pregnancy Rate of preterm delivery/1000 deliveries Primigravida Multigravida Age at marriage ≤14 y 220 186 15 y 197 193 16–17 y 191 196 ≥18 y 165 190 Age at first pregnancy ≤15 y 226 199 16–17 y 193 196 ≥18 y 202 190 Note: Rates of preterm delivery per 1000 live deliveries for each category for age at marriage and age at first pregnancy. Abbreviation: y, years of age.

Adjusted odds ratios (aOR) for the association of age at marriage with preterm delivery. Association between preterm delivery and age at marriage for (A) primigravida participants and (B) multigravida participants. Core confounders identified using a directed acyclic graph were maternal caste, maternal education, and household asset score. Minimally adjusted: Adjusted only for random effect of cluster and fixed effect of study arm and strata, using multivariate imputation by chained equations. Fully adjusted: Adjusted for cluster, study arm, strata and core confounders, using multivariate imputation by chained equations. Fully adjusted plus age at pregnancy: Adjusted for cluster, study arm and strata and core confounders + age at pregnancy, using multivariate imputation by chained equations. Abbreviations: aOR, adjusted odds ratio; y, years of age. p-value significance: * <.05, ***<.01

Figure 4 displays the aOR for the association between age at first pregnancy and preterm delivery. Primigravida participants giving birth aged ≤15 years had a nonsignificant higher odds of preterm delivery in the minimally adjusted model (aOR 1.18, 95% CI: 0.94, 1.48), however this association was substantially attenuated in fully adjusted models. No significant associations were observed between age at first pregnancy and preterm delivery in multigravida participants in any model (Figure 4).

Adjusted odds ratios (aOR) for the association of age at first pregnancy with preterm delivery. Association between preterm delivery and age at first pregnancy for (A) primigravida participants and (B) multigravida participants. Core confounders identified using a directed acyclic graph were maternal caste, maternal education, household asset score and age at marriage. Minimally adjusted: Adjusted only for random effect of cluster and fixed effect of study arm and strata, using multiple imputation by chained equations. Fully adjusted: Adjusted for cluster, study arm, strata and confounders, using multivariate imputation by chained equations. Fully adjusted minus age at marriage: Adjusted for cluster, study arm, strata and confounders (minus age at marriage), using multivariate imputation by chained equations. Abbreviations: aOR, adjusted odds ratio; y, years of age. p-value significance: *<.05, ***<.01

Supplemental Table 3 displays investigations into whether markers of nutritional status explain our observed associations. While models including BMI and MUAC were limited by a high amount of missing data, a significant increase in the odds of delivering preterm for primigravida participants married <14 years was m