Relationship between early-life nutrition and ages at menarche and first pregnancy, and childbirth rates of young adults: Evidence from APCAPS in India.

India's Integrated Child Development Services (ICDS) provides daily supplementary nutrition and other public health services to women and children. We estimated associations between exposure to early-childhood ICDS nutrition and adult reproductive outcomes. During 1987-1990, a balanced protein-calorie supplement called "upma"-made from locally available corn-soya ingredients-was rolled out by subdistricts near Hyderabad and offered to pregnant women and children under age 6 years. In a controlled trial, 15 villages received the supplement and 14 did not. We used data from a 2010-2012 resurvey of adults born during the trial (n = 715 in intervention and n = 645 in control arms). We used propensity score matching methods to estimate the associations between birth in an intervention village and menarcheal age, age at first pregnancy, and fertility of adults. We found that women born in the intervention group during the trial, as compared with the control group, had menarche 0.45 (95% confidence interval [CI: 0.22, 0.68]; p < .001) years later and first pregnancy 0.53 (95% CI [0.04, 1.02]; p < .05) years later. Married women from the intervention group had menarche 0.36 (95% CI [0.09, 0.64]; p < .01) years later, first cohabitation with partner 0.8 (95% CI [0.27, 1.33]; p < .01) years later, and first pregnancy 0.53 (95% CI [0.04, 1.02]; p < .05) years later than married women in the control group. There was no significant difference between intervention and control group women regarding whether they had at least one childbirth or the total number of children born. The findings were similar when we employed inverse propensity score weighted regression models.


| INTRODUCTION
Progress on child survival is among the most significant social achievements in India during the 21st century to date. Under-5 child mortality in India has fallen sharply from 92 per 1,000 live births in 2000 to 39 in 2017 (World Bank, 2017). The rate of decline has accelerated over time, leading to nearly a million more avoided under-5 deaths during 2005-2015 when compared with (Fadel et al., 2017.
However, rates of undernutrition among Indian children remain high.
In 2016, 38%, 21%, and 36% of Indian children under 5 were stunted, wasted, and underweight, respectively (International Institute for Population Sciences [IIPS], 2016). In 48 districts of India-primarily from poorer and more populous states-more than half of under-5 children were stunted, and in 38 districts, more than half were underweight (IIPS, 2016).
Little is known about how early-life supplementary nutrition affects the reproductive health of LMIC adults, including ages at menarche, first cohabitation with partner, and first pregnancy, and the rate and number of childbirths. Observational studies in India, Peru, Vietnam, Kenya, Senegal, and Guatemala have found associations of higher birthweights and lower prepubertal body mass with later menarche and greater height for age at 24 months with lower adult fertility (Aurino, Schott, Penny, & Behrman, 2017;Hoddinott, Behrman, et al., 2013b;Khan, Schroeder, Martorell, Haas, & Rivera, 1996;Leenstra et al., 2005;Simondon et al., 1998). Although several studies in highincome countries have found that later menarche is associated with greater adolescent and adult height, such relationships have not been examined in LMICs (Aurino et al., 2017;Frisch & Revelle, 1970;He & Karlberg, 2001;Zacharias & Rand, 1983).
In this study, we followed up 1,360 adults of age 20-25 years who were born during the original APCAPS nutritional trial. We estimated associations between exposure to supplementary nutrition and adult reproductive outcomes.

Key messages
• Adequate nutrition during the first 2-3 years of life can have lasting positive impact on health through adulthood.
• The relationship between early-life nutrition and adult reproductive outcomes are incompletely understood, especially in low-and middle-income countries.
• This study found supplementary nutrition in-utero and during the first 3 years of life to be longitudinally associated with delayed menarche and delayed first pregnancy among Indian women • The findings have important policy implications for maternal and child health.

| APCAPS data
India's Integrated Child Development Services (ICDS) programme is the largest public supplementary nutrition and development programme for women and young children in the world. It provides a daily cooked meal or uncooked take-home dry meal to pregnant and nursing women, children below the age of 6 years, and adolescent girls. Additional services under the ICDS include preschool education and immunization for children, health check-ups and referrals, and nutrition and health education In the intervention villages, all pregnant women and children below the age of 6 years were offered daily cooked meals at the Anganwadi centres, which they could eat there or take home. The meals-called "upma"-were prepared from locally available corn-soya blends and provided on average 500 kcal of energy and 20-25 g of protein to women and half of those amounts to children (Kinra et al., 2008). In control villages, supplementary nutrition was introduced 3 years after the study ended. Prior to the trial, the intervention and control groups had similar availability of public health services such as an anaemia control in pregnancy, routine childhood vaccination, and primary health care. An abstract describing the trial was published, but no further data-including information on consumption of cooked meals by participants-are available (Kinra et al., 2008).
The study enrolled all children born between January 1, 1987, and Together, the surveys have collected data on a wide range of outcomes including socio-economic and demographic factors, anthropometry, lifestyle, medical history, cardiovascular physiology, spirometry, and biomarkers . The original trial and the followup surveys are collectively known as the APCAPS. The data collection processes have been discussed in previous studies (Kinra et al., 2008Kinra, Sarma, Hards, Smith, & Ben-Shlomo, 2011).

| Ethics statement
Ethical clearance for the APCAPS cohort study was provided by the

| Propensity score matching analysis
We used the 2010-2012 wave of the APCAPS survey, in which those who were born during the original trial (hereafter referred to as index adults) were 20-25 years old. We estimated associations between birth in an intervention village and the following reproductive outcomes of index adults: (a) ages at menarche in years for women, (b) ages at first pregnancy in years for women, (c) binary indicator of whether the adults had at least one child, and (d) total number of children (including zero) born to the adults. All outcomes were selfreported.
As discussed in previous APCAPS studies, there were some statistically significant differences in socio-economic and demographic characteristics of the intervention and control group adults, which could bias ordinary least square estimates of associations between birth in intervention villages and outcome variables Nandi et al., 2016;. We used a quasi-experimental propensity score matching (PSM) technique to mitigate such differences (Dehejia & Wahba, 2002;Heckman, Ichimura, & Todd, 1997;Leuven & Sianesi, 2003;Rosenbaum & Rubin, 1984).
First, we conducted a probit regression of birth in intervention villages for index adults on household and individual socio-economic and demographic characteristics determined before the trial including age FIGURE 1 Participants of the APCAPS in years, sex (whether female), indicators of caste groups (scheduled caste or scheduled tribe, and other backward classes) and religion (non-Hindu), and parental schooling attainment (literate, completed primary, and completed secondary or above).
Based on the predicted probability from the probit model (known as the propensity score), we matched intervention group index adults with similar index adults from the control group. We restricted our sample to individuals with overlapping propensity scores from the two groups (known as "common support") and used a Kernel (Epanechnikov) matching algorithm (Caliendo & Kopeinig, 2008;Dehejia & Wahba, 1999;2002).
Our estimation approach assumes that unobserved characteristics of the intervention and control groups were randomly distributed after matching. Consequently, the average difference in an outcome variable between the intervention and matched control group could be attributed to supplementary ICDS nutrition. The difference is known as the "intent-to-treat" estimator.
In addition to the analyses of fertility outcomes of all index adults, we also evaluated outcomes separately for women and ever-married women. We considered p < .05 for statistical significance.

| Inverse propensity score weighted regression analysis
We used additional propensity score weighted multivariate regression models to examine associations of the intervention (Hirano & Imbens, 2001;Hirano, Imbens, & Ridder, 2003;McCaffrey, Ridgeway, & Morral, 2004). This technique balances the distribution of covariates between the intervention and control groups and produces statistically efficient estimators. We used the propensity score derived from the first-stage probit model of PSM as discussed earlier. The weight was defined as the inverse of the propensity score for the intervention group and inverse of one minus the propensity score for the control group (Hirano et al., 2003;Hirano & Imbens, 2001;McCaffrey et al., 2004).
These weights were employed in regression analyses of outcomes.
We used Cox proportional hazards models for age at menarche and age at first pregnancy, probit regression for the binary indicator of whether the adult had at least one child, and Poisson regression for total number of children (including zero) born to the adult.
Along with the indicator of birth in intervention villages, the following variables were included as regression covariates: age in years and sex (whether female) when applicable, indicators of caste groups and religion, and parental schooling attainment. Standard errors were robust clustered at the village level.

| Tests of matching quality and additional sensitivity analyses
We conducted three tests of matching quality in the base model, that is, whether PSM was successful in reducing differences in characteristics between the intervention and control groups. We computed the standardized percentage bias-which was the difference of the sample means of a covariate between the two groups as a percentage of the square root of the average of the sample variances of the groupsseparately before and after matching (Rosenbaum & Rubin, 1985). A successful PSM would reduce the average bias substantially from prematching to postmatching. We also re-estimated the propensity score regression using the matched intervention-control sample.
Under a valid PSM, the pseudo-R 2 of this model should be substantially lower as compared with the original propensity score model (Sianesi, 2004). Finally, we conducted a likelihood ratio test of the joint significance of covariates prematching and postmatching. Failure to reject the null hypothesis of joint significance after matching would indicate that the matched sample was homogenous in nature and that the covariates would not adequately explain intervention status any more after matching.
Age at first pregnancy may be correlated with age of marriage (Nandi et al., 2018). Any associations of the intervention with this earlier-in-life event could possibly drive its association with ages at first pregnancy. We conducted additional analyses to explore these potential pathways, as discussed in the Supporting Information. We also tested for sensitivity of PSM by examining additional matching algorithms, as presented in the Supporting Information.

| Characteristics of the study sample
Summary statistics of all index men and women are presented in Table 1. In the intervention and control groups, there were 715 and 645 index adults (of them, 272 and 252 women), respectively. There were 160 and 167 married women, respectively, in the two groups, and their summary statistics are presented in Table S5. There was no statistical difference in the proportion of married women between the two groups. Due to missing covariate values, two sample observations were lost for analysis.
Intervention village adults had completed secondary education and were employed or enrolled in higher education at significantly higher proportions. Lower proportions of them were ever married, and they had fewer children than control group adults. Intervention group women had higher menarcheal and first pregnancy ages as compared with control group women. The proportion of intervention group households belonging to scheduled caste or scheduled tribe groups was higher compared with the control group, whereas those belonging to other backward classes was lower. Higher proportions of intervention group households belonged to the middle three wealth quintiles as compared with the control group.

| Propensity score matching estimates
The PSM intent-to-treat estimators of associations between birth in an intervention village and adult reproductive outcomes are presented in Table 2. Note. Values are mean ± standard deviation, unless stated otherwise. Difference is between the mean values of intervention and control groups. Index adults are those born in study villages during the original trial period of 1987 to 1990 and alive at the time of the survey. Wald t tests for continuous variables and z tests for proportions were used to examine the statistical significance of the differences between intervention and control group means.

| Results from inverse propensity score weighted regression
Results of weighted regression analysis are presented in Table 3. Menarcheal ages were higher among intervention village women and mar- p < .05) fewer children as compared with control group women.

| Results from tests of matching quality
As shown in Table 4, PSM successfully reduced differences in the characteristics of the intervention and control groups. The mean standardized percentage bias, that is, average difference between the two groups in the covariates of the first stage regression, reduced by 77-90% from prematching to postmatching across all models. Only in the subsample of married index men, the reduction was smaller, likely due to low sample size. Detailed results of these balance tests for each covariate are presented in Tables S6 and S7.
Similarly, the value of pseudo-R 2 reduced by more than 90% from prematching to postmatching, and the likelihood ratio test statistic of joint significance of the covariates became statistically significant to nonsignificant from prematching to postmatching. These results show that the PSM method was valid.

| DISCUSSION
Using PSM and weighted regression analysis, we observed that 20to 25-year-old women who were exposed to supplementary nutrition in utero and during the first 3 years of life under a controlled trial had menarche and first cohabitation with a partner at older ages as compared with observationally identical women who were not exposed to the supplementary nutrition. Intervention group women (and men) had fewer childbirths compared with matched control group women (and men), although no difference was seen among married women (and men). Because there were no childbirths reported by unmarried adults, differences in childbirths between intervention and control groups were explained fully by lower  Note. Values are the estimated coefficient of the indicator of birth in an intervention village in the regression of the outcome variable, along with 95% CIs, unless stated otherwise. Estimated coefficients of other regression covariates are not shown. We used Cox proportional hazards model for age at menarche and age at first pregnancy, marginal effect probit regression models for whether the adult had at least one childbirth, and Poisson regression models for the number of children born. For Cox models, hazard ratios are shown instead of coefficients. All regression models were weighted using inverse propensity score weights. Standard errors were clustered at the village level.
proportions of marriage in the intervention group (Nandi et al., 2018). In PSM analysis, but not in weighted regressions, intervention group women were also observed to have later first pregnancies as compared with similar control group women. The results were not sensitive to matching algorithms.
The determinants of menarcheal age in LMICs have not been widely studied. A recent study using longitudinal Young Lives data on adolescents in India, Peru, and Vietnam found positive associations of birthweight and negative associations of prepubertal height and body mass measures with menarcheal ages (Aurino et al., 2017). Prepubertal stunting was linked with later menarche also in Guatemala, Senegal, and Bangladesh (Bosch, Willekens, Baqui, Van Ginneken, & Hutter, 2008;Khan et al., 1996;Simondon et al., 1998).
In our study, exposure to early-life nutrition was associated with later menarche and higher variances in menarcheal age. Due to lack of data on birthweight or prepubertal body mass of study participants, we could not examine their relationships with menarcheal age. However, our findings align with previous studies linking earlier puberty with earlier cessation of growth in adolescents (Sandhu et al., 2006;Zacharias & Rand, 1983). The average menarcheal age in our study was 12.8 years, similar to the average in Andhra Pradesh (95% CI [12.7, 13.1 years]) at that time. Then at ages 13-18 years, intervention group participants were 1.4 cm taller (Kinra et al., 2008), possibly due to earlier menarche and growth cessation in the control group. A study of 1,520 men in the United Kingdom found that adolescents with later puberty were 0.6 cm taller as compared with those with earlier puberty (Sandhu et al., 2006). Another study of 286,000 adult women from 10 European countries associated a 1-year later menarche with 0.3-cm growth (Onland-Moret et al., 2005).
Soy protein-an ingredient of the meals offered to the intervention group-may have also affected menarcheal age in our study. Previous studies of adolescents in Germany and the United States have found that higher levels of vegetable protein intake, as compared with lower intake, could delay menarche (Berkey, Gardner, Frazier, & Colditz, 2000;Günther, Karaolis-Danckert, Kroke, Remer, & Buyken, 2010;Kissinger & Sanchez, 1987). Naturally occurring endocrine disruptors in soy could also delay puberty in girls, although the evidence is not conclusive (Fisher & Eugster, 2014). Note. Standardized percentage bias was measured as the difference of the sample means of a covariate between the two groups as a percentage of the square root of the average of the sample variances of the groups. Matching was based on propensity scores, using the Kernel (Epanechnikov) matching method.
The association between early-life nutrition and age at first pregnancy could be mediated by schooling attainment and age at marriage.
Women who completed schooling earlier than other women may be more likely to get married and in turn more likely to become pregnant earlier. In a previous study of the APCAPS adult cohort, we found that higher proportions of intervention group women completed secondary-and graduate-level schooling and lower proportions were married, as compared the control group (Nandi et al., 2018). In the current study, ages at first cohabitation were higher in the intervention group than in the matched control group. This indicates that the pathway of positive association between nutrition and first pregnancy age may be through later cohabitation in the intervention group. In the INCAP study, a one standard deviation increase in height-for-age z scores at age 2 years was associated with 0.77 years higher ages at first childbirth, 0.63 fewer pregnancies, and 0.43 fewer childbirths among 25-to 42-year-old women (Hoddinott, Behrman, et al., 2013c). No associations were seen between height-for-age z scores and age at first marriage.
Our findings have important policy implications. The importance in LMICs of adult height as an indicator of past nutrition and as a determinant of adult health status, cognitive ability, schooling attainment, and economic productivity is well accepted (Almond & Currie, 2011;Currie & Vogl, 2013;Victora et al., 2008). Therefore, supplementary nutrition, which may delay menarche, could potentially improve adult height and cognitive, schooling, and economic outcomes (Aurino et al., 2017).
Ages at first cohabitation and first pregnancy have important implications for maternal and child health and well-being (Gibbs et al., 2012;Yu, Mason, Crum, Cappa, & Hotchkiss, 2016). In India, a 1-year delay in marriage for women is associated with lower home birth and higher breastfeeding rates and improved vaccination rates, weight-forheight, school enrolment, and cognitive test scores of children (Chari et al., 2017). In a study of five LMICs including India, children of mothers below the age of 19 years were more likely to be born preterm, had lower birthweights and higher rates of 2-year stunting, and were less likely to complete secondary schooling than children of 20-to 24-year-old mothers (Fall et al., 2015). In our data, 43% and 38% of married index women started cohabiting with partners and had their first children, respectively, before the age of 19 years.
At the national level, 27% of 20-to 24-year-old Indian women were married before the age of 18, and 8% of 15-to 19-year-old women in 2016 had pregnancies or childbirths (IIPS, 2016), possibly leading to substantial adverse health effects for their children.
There are limitations to our analysis. Access to public health services such as anaemia control in pregnancy and routine child immunization was similar across intervention and control villages before the trial. Integration of these services though the introduction of daily meals may have increased their use in the intervention villages, on which no data are available. However, such services may increase effects of nutrition by improving overall health of index children and their mothers.
Reproductive outcomes in our study were self-reported by participants. Data on age at menarche were collected in adolescence, limiting measurement error arising from a long recall period. Data were collected by female surveyors only to limit reporting bias. The average age at menarche in our study was consistent with other concurrent sources from the state (Pathak et al., 2014). Age at first pregnancy and fertility are less likely to be affected by the duration of recall period. Whereas random error, which could underestimate estimated associations, cannot be entirely ruled out, systematic (recall) bias is unlikely as the study participants were unaware of these secondary hypotheses at the time of data collection.
No information was available on which study participants actually consumed the meals offered to them during the trial. Any unknown redistribution of meals within or across households might have weakened the estimated associations with reproductive outcomes in our study. There was also no information on future marriage and pregnancies of unmarried women, requiring us to deal with censored data.

ICDS supplementary nutrition was introduced in control villages
after the trial ended. Index adults in these villages were exposed to nutrition during 3-6 years of their lives, potentially weakening the estimated nutritional associations in our study. However, nutrition during the first 2-3 years of life is known to have the largest effect on most future outcomes (Black et al., 2013;Grantham-McGregor et al., 2007).

| CONCLUSION
Exposure to supplementary nutrition in utero and during the first 3 years of life was associated with higher ages at menarche, first cohabitation with partners, and first pregnancies among Indian women.

ACKNOWLEDGMENTS
The authors are grateful to the sponsors, fieldworkers, and partici- provided support in kind to the follow-up surveys through free or subsidized access to facilities and materials. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

CONFLICTS OF INTEREST
The authors declare that they have no conflicts of interest.

CONTRIBUTIONS
SK and other collaborators at APCAPS collected the data. AN conducted the analysis and wrote the first version of the manuscript.
JRB, SK, MMB, and RL interpreted the findings and edited the manuscript. AN, RL, and JRB were responsible for the final contents. All authors have read and approved the final manuscript.