Early Childhood Nutrition Is Positively Associated with Adolescent Educational Outcomes: Evidence from the Andhra Pradesh Child and Parents Study (APCAPS)^,,

Abstract

Background: India’s Integrated Child Development Scheme, which provides supplementary nutrition and other public health services to >91 million women and children aged <6 y, is the largest program of its kind in the world.

Objective: We estimated the long-term associations of maternal and early childhood nutrition provided under the Integrated Child Development Scheme with educational outcomes when the children became adolescents.

Methods: We used longitudinal data from a controlled nutrition trial conducted near the city of Hyderabad, India. From 1987 to 1990, a balanced protein-energy supplement (corn-soya meal, called upma) was offered to pregnant women and children aged <6 y in 15 intervention villages, whereas no supplementation was offered in 14 control villages. Both groups had equal access to other public programs such as immunization and anemia control in pregnancy. Children born during the original trial period were resurveyed (654 intervention and 511 control group children) in 2003–2005. We used propensity score matching methods to correct for estimation bias in our regression models to assess the associations of supplementary nutrition with school enrollment, schooling grades completed, and academic test performance of these adolescents.

Results: Children born in intervention villages were 7.8% (95% CI: 0.1%, 15.4%; P < 0.05) more likely to be enrolled in school and completed 0.84 (95% CI: 0.28, 1.39; P < 0.005) more schooling grades than children born in control villages. We found no association between supplementary nutrition and academic performance, as measured by school test scores.

Conclusion: Offering a nutritional supplement to pregnant women and children <6 y of age during the Hyderabad Nutrition Trial was associated with improved school enrollment and completion of more schooling grades when the children became adolescents.

Introduction

An estimated 200 million children <5 y of age—most of whom live in South Asia and sub-Saharan Africa—fail to achieve their potential primarily because of undernutrition (1–3). Undernutrition during the first 1000 d of life is associated with lower survival, higher incidence of acute and chronic diseases in later life (3, 4), reduced cognitive skills, fewer schooling grades (5–7), lower economic productivity (8), and lower birth weight and shorter children in the next generation (9). Consequently, undernutrition among pregnant women and young children has drawn substantial attention from researchers and policy makers (8, 10, 11).

Adequate nutrition in early life may help break the cycle of poor outcomes from childhood to adulthood. Under the “fetal origins” hypothesis, better nutrition in utero can significantly improve later-life health and educational and economic outcomes (4, 12). However, data supporting the hypothesis are primarily from high-income countries and typically use birth weight to measure nutrition in utero rather than through interventions that improve maternal nutrition (4, 5, 13, 14). Also, a number of these studies examine the effects of nutrition in utero or during early childhood only on future health outcomes (4).

Long-term educational or economic benefits of improving nutrition in utero or in the early years of life are poorly quantified among children in low- and middle-income countries (LMICs)⁹. A systematic review found that the relation between early childhood nutrition and adult outcomes had been researched so far in just 8 developing countries (14). Seven of these country studies used observational data and found positive associations primarily with health outcomes such as height and weight, disability, and prevalence of heart diseases. Another review study (5) analyzed the Consortium of Health-Orientated Research in Transitioning Societies (15) data from Brazil, Guatemala, India, the Philippines, and South Africa and found a knowledge gap especially on the association between maternal nutrition and children’s educational outcomes. Little is known about how maternal height (a general measure of mothers’ lifetime nutrition) or supplemental nutrition during pregnancy might affect child outcomes in LMICs (5).

Our understanding of the long-term consequences of early childhood nutrition comes from a few experimental studies with longitudinal follow-up. The Institute of Nutrition for Central America and Panama study that followed ~1500 children in 4 villages of Guatemala between 1969 and 2004 found that protein-rich nutritional supplementation before 24 mo of age was associated with 1.2 higher schooling grades completed for women, increased adult cognitive skills for both men and women by around one-fourth of an SD, and increased average adult male wages by 46% (16–18). A study in which 390 adolescent Nepalese children were randomly allocated to maternal vitamin A supplementation and a placebo control found that the intervention was associated with an 11.3–percentage point reduction in schooling repetition rates but no association with test scores (19). Another study evaluated exposure to iodine supplementation in utero among 1400 adolescent Tanzanian children. Using a quasi-experimental method involving variations in district-level coverage rates of an iodine program, the authors found that children exposed to the supplementation completed 0.4–0.6 more schooling grades, with larger effect sizes for girls (20). In a cohort study involving 3493 participants in Brazil, at least 12 mo of breastfeeding in childhood was associated with a 3.76-point increase in intelligence quotient, 0.9 increase in grades completed, and 341 Brazilian Reals (approximately US$168) increase in income at age 30 y (21). Finally, Muslim Indonesian children exposed to in utero daylight fasting during Ramadan do more child labor; score 7.4% and 8.4% lower on cognitive and math tests, respectively; and, as adults, work 4.7% fewer hours and are more likely to be self-employed (22).

Methods

Data: Integrated Child Development Scheme and the Hyderabad Nutrition Trial.

The Integrated Child Development Scheme (ICDS) program of India, launched in 1975 and universalized in 2008–2009, is the world’s largest maternal and child health program (23). It is designed to address nutritional and developmental needs of children aged <6 y, pregnant women, and nursing mothers. The ICDS has a network of >1.3 million mother-child health centers, known as Anganwadi, that provide supplemental nutrition, preschool education, immunization, health checkups and referral services, and nutrition and health education (23, 24). In 2011, the program provided nutrition to 91.9 million women and children, as well as preschool education to 35.5 million children (25).

The 1987–1990 Hyderabad Nutrition Trial, conducted by the National Institute of Nutrition, evaluated effects of food supplementation during pregnancy under the ICDS on newborn birth weights. The study enrolled pregnant women and young children from 29 villages in administrative areas near the city of Hyderabad in the state of Andhra Pradesh (now part of the new state of Telangana). First, 2 adjacent subdistricts (group of villages, called a block) were selected for the study: one for the intervention group and the other as a control. Then, a prominent village at the center of the block was randomly located on the map, and all villages within a 10-km radius were selected for the study, resulting in 15 intervention and 14 control villages. Expectant mothers and children up to age 6 y in the intervention villages received daily food supplements under the ICDS—a corn-soya meal (upma) made using locally available ingredients—as part of the study. The meal provided a mean of ∼2.09 MJ (500 kcal) of energy and 20–25 g of protein to pregnant or lactating women and ∼1.25 MJ (300 kcal) of energy and 8–10 g of protein to children (26, 27). The supplement was offered in intervention villages during 1987–1990, and the program was expanded to include the control villages 3 y after the end of the study. Intervention and control group villages had similar levels of access to other public health programs such as immunization and anemia control in pregnancy. A short abstract describing the study was published, but no further study documents or reports are available (26).

In 2003–2005, households with at least one living child born during the original trial period were located and invited to take part in a follow-up study (along with further follow-ups; it is now known as the Andhra Pradesh Child and Parents Study, or APCAPS) (28). Complete details of the trial and follow-up study were published in Kinra et al. (26) and are not repeated here. A brief summary is presented below.

The follow-up survey first identified women (from 1815 families) who participated in the original trial. These women reported a total of 2601 children born during 1987–1990 who were still alive in 2003. These children were then matched with baseline data using date of birth and sex information. The study team set up a health clinic in each village and invited children who were successfully matched with baseline for the follow-up survey at the clinic. A total of 1165 such children—654 in intervention villages and 511 in control villages—attended the clinics (see Figure 1). The survey collected various data—including height, adiposity, blood pressure, and insulin resistance—for the then 13- to 18-y-olds. Self-reported information on current school enrollment status, highest schooling grade completed, and performance on the last major test at school for each child was also collected (28). Furthermore, household and parental socioeconomic, demographic, and educational data were gathered. Summary statistics of the follow-up study sample are presented in Table 1.

FIGURE 1.

Number of children in the intervention and control groups of the 2003–2005 follow-up of the Hyderabad Nutrition Trial. Adapted from reference 26 with permission.

TABLE 1.

Demographic and socioeconomic characteristics of children in the 2003–2005 follow-up of the Hyderabad Nutrition Trial¹

	Intervention group	Control group
Child characteristics
Born in 1987	0.291 ± 0.455 (184)	0.289 ± 0.454 (138)
Born in 1988	0.443 ± 0.497 (280)	0.397 ± 0.49 (190)
Born in 1989	0.179 ± 0.383 (113)	0.236 ± 0.425 (113)
Born in 1990	0.087 ± 0.282 (55)	0.077 ± 0.268 (37)
Birth order	2.922 ± 1.502 (632)	3.086 ± 1.689 (478)
Male child	0.531 ± 0.499 (336)	0.538 ± 0.499 (257)
Social characteristics of household
Scheduled caste/scheduled tribe	0.384 ± 0.487 (243)	0.291 ± 0.455 (139)
Other backward caste	0.486 ± 0.5 (307)	0.569 ± 0.496 (272)
Religion other than Hindu	0.033 ± 0.179 (20)	0.059 ± 0.235 (28)
Standard of living
Wealth quintile 1	0.228 ± 0.42 (144)	0.262 ± 0.44 (125)
Wealth quintile 2	0.162 ± 0.368 (102)	0.155 ± 0.362 (74)
Wealth quintile 3	0.231 ± 0.422 (146)	0.231 ± 0.422 (110)
Wealth quintile 4	0.182 ± 0.386 (115)	0.195 ± 0.397 (93)
Wealth quintile 5	0.197 ± 0.398 (124)	0.157 ± 0.364 (75)
Father’s education
Literate	0.159 ± 0.366 (100)	0.139 ± 0.346 (66)
Primary	0.113 ± 0.316 (71)	0.151 ± 0.359 (72)
Secondary and above	0.098 ± 0.298 (62)	0.126 ± 0.332 (60)
Mother’s education
Literate	0.043 ± 0.203 (27)	0.038 ± 0.191 (18)
Primary and above2	0.05 ± 0.217 (31)	0.076 ± 0.265 (36)

¹Values are mean proportion ± SD (n). All variables except birth order are binary. Intervention group includes children born during 1987–1990 in villages where the nutritional supplement was provided. Control group includes children born during 1987–1990 in the control villages.

²Because very few mothers have a secondary or higher level of education, we combined them with the category “primary and above.”

Ethics.

Ethical approval for the study was obtained from the Ethics Committee of the National Institute of Nutrition of India, Hyderabad. Approval was also obtained from the local administrative authorities in all villages. Informed consent was obtained from the participants and their parents or guardians.

Estimation procedures.

We used the 2003–2005 follow-up data to estimate the differences in educational outcomes between children born in intervention and control areas of the trial. Three educational outcomes were examined: 1) enrollment (whether children were currently enrolled in school), 2) schooling grade (0–12) completed by children, and 3) academic performance (most recent test scores at school). Test scores were measured as follows: fail = 1, third class = 2, second class = 3, first class = 4, and distinction = 5. We used discrete-choice econometric models for analyzing these educational outcomes: probit for current school enrollment and ordered probit models for schooling grade completed and academic performance. We included both currently and previously enrolled children in our analysis. Therefore, the schooling grade completed by a child captured a combined effect of age of school entry, schooling grade repetition, and possible dropout. Data were analyzed with the use of STATA version 12 (StataCorp LP).

Our regression models controlled for children’s age, sex, birth order, and father’s and mother’s education (literate, completed primary, and completed secondary or above). Also included were household socioeconomic characteristics, including caste (scheduled caste or scheduled tribe, other backward class), religion (Hindu or other), and household assets (26, 29, 30). SEs were clustered at the village level through the use of the cluster() option in STATA regression commands.

Among children invited to participate in the 2003–2005 follow-up survey, 82% in the intervention villages and 74% in the control villages complied. The participants represented 49% and 41%, respectively, of all children born in the study villages during 1987–1990 who were alive in 2003 (see Figure 1). To mitigate effects of possible systematic differences between intervention and control children due to selective attrition or other reasons, we conducted additional analyses by means of the quasi-experimental method of propensity score matching (PSM) (31–35). PSM compares each individual affected by the intervention with a control individual(s) who was not affected by the intervention but had basically the same probability of being affected by the intervention on the basis of observed characteristics. This comparison attributes the differences in educational outcomes solely to the intervention if any unobservable factors that affect the intervention are distributed randomly between intervention and control groups.

In the PSM analysis, we regressed intervention assignments on children’s background characteristics. Then, based on predicted probabilities (propensity scores) from the estimated model, we matched intervention group children with similar control group children. We used the one-to-one nearest-neighbor matching method (with replacement) within a probability radius of 0.01. We considered only observations from the 2 groups that were in “common support”—that is, with overlapping propensity scores. SEs of our analysis were adjusted by using a bias correction method (36). Further methodologic details of the PSM analysis are presented in the Supplemental Methods.

After matching, the differences in mean educational outcomes between the intervention and matched control children yield the estimated nutritional supplement associations (termed the mean treatment effect on the treated). However, the estimated associations can vary depending on the choice of matching method. We therefore examined the robustness of our findings by considering additional propensity score estimation techniques and matching algorithms. We tested several alternative strategies: a kernel (Epanechnikov) method for matching intervention and control children instead of a nearest-neighbor match, matching each intervention child with 5 nearest neighbors in the control group instead of a single neighbor, and using a logit model of propensity score estimation (nearest-neighbor matching with replacement within a radius of 0.01) instead of a probit model. Finally, we also used a covariate matching technique that matched an intervention observation with its nearest control group counterpart on the basis of the joint distribution of all covariates, instead of reducing the dimension to a single propensity score (37). Results from our regression and PSM analyses are presented as percentage change in the mean current school enrollment rate and changes in the mean values of schooling grade completed and test scores at school. Estimates are considered statistically significant at the 5% level.

It is also important to ensure that the covariates were balanced and systematic differences between intervention and control groups were reduced after PSM. We evaluated this in 3 different ways. First, we calculated the percentage bias (difference) in mean values of covariates between the intervention and unmatched control groups. We then examined whether the percentage bias declined significantly if the intervention group was instead compared with the matched control group. Second, we tested matching quality by using the pseudo-R² statistic from the reestimation of propensity scores from the matched sample, following Sianesi (38). If matching quality is good, pseudo-R² should decrease substantially from the unmatched to the matched data. Third, a likelihood ratio test of joint significance of the covariates was conducted separately in the unmatched and matched data. A failure to reject the null hypothesis of joint nonsignificance in the matched data indicates good matching quality.

Results

Probit results are reported in Table 2. The first row shows estimated associations of the intervention with educational outcomes. Controlling for child, parental, and household characteristics, the intervention was associated with a 6.6% (95% CI: 0.8%, 12.4%; P < 0.05) rise in current school enrollment. We did not find any significant association between the intervention and schooling grade completed or test scores.

TABLE 2.

Probit and ordered probit regressions of educational outcomes of children in the 2003–2005 follow-up of the Hyderabad Nutrition Trial¹

	School enrollment	Schooling grade completed	Academic performance
Born in an intervention village	0.066 (0.008, 0.124)	0.128 (−0.122, 0.379)	0.055 (−0.077, 0.188)
Age	−0.083 (−0.111, −0.056)	0.335 (0.256, 0.415)	−0.013 (−0.096, 0.069)
Birth order	−0.001 (−0.015, 0.012)	−0.022 (−0.067, 0.024)	0.009 (−0.035, 0.053)
Male child	0.217 (0.165, 0.268)	0.378 (0.231, 0.526)	0.344 (0.148, 0.54)
Scheduled caste/scheduled tribe	−0.036 (−0.128, 0.056)	−0.449 (−0.799, −0.099)	−0.267 (−0.522, −0.012)
Other backward caste	−0.093 (−0.174, −0.011)	−0.326 (−0.617, −0.035)	−0.252 (−0.508, 0.003)
Non-Hindu household	−0.400 (−0.656, −0.144)	−0.205 (−0.675, 0.265)	−0.367 (−0.759, 0.025)
Wealth quintile 2	0.056 (0.009, 0.102)	0.180 (0.033, 0.326)	−0.125 (−0.33, 0.08)
Wealth quintile 3	0.062 (0.029, 0.095)	0.099 (−0.049, 0.247)	−0.137 (−0.346, 0.071)
Wealth quintile 4	0.066 (0.018, 0.113)	0.400 (0.148, 0.652)	−0.29 (−0.496, −0.085)
Wealth quintile 5	0.091 (0.037, 0.145)	0.242 (0.036, 0.449)	0.295 (0.03, 0.559)
Father literate	0.107 (0.06, 0.153)	0.194 (0.032, 0.356)	0.136 (−0.062, 0.334)
Father’s education: primary	0.047 (−0.001, 0.096)	0.210 (0.046, 0.374)	0.282 (−0.008, 0.573)
Father’s education: secondary and above	0.140 (0.101, 0.178)	0.529 (0.23, 0.828)	0.186 (−0.077, 0.449)
Mother literate	0.130 (0.082, 0.178)	0.369 (0.032, 0.706)	0.034 (−0.262, 0.33)
Mother’s education: primary and above2	0.116 (0.04, 0.192)	0.627 (0.303, 0.951)	0.367 (0.041, 0.693)
P value for χ2	0.000	0.000	0.000
Sample size, n	1096	1059	904

¹Values are regression coefficients with 95% CIs in parentheses, unless stated otherwise. The number of observations differs across regression models because of some missing values of the outcome variables.

²Because very few mothers have a secondary or higher level of education, we combined them with the category “primary and above.”

Differences among intervention, unmatched control, and matched control observations are presented in Figure 2. We found that intervention group children were on average 6.9% (95% CI: 2.1%, 11.6%; P < 0.01) more likely to be currently enrolled in school (Figure 2A) and completed 0.5 (95% CI: 0.3, 0.8; P < 0.001) additional schooling grades (Figure 2B) compared with children in the unmatched control group. There was no statistically significant difference in their test scores (Figure 2C).

FIGURE 2.

Mean educational outcomes of children in the 2003–2005 follow-up of the Hyderabad Nutrition Trial—intervention, unmatched control, and matched control groups. The matched control observations were obtained from propensity score matching by using a one-to-one nearest-neighbor match with replacement within a radius of 0.01. Current enrollment (A) is binary, whereas schooling grade completed (B) and test score at school (C) are measured on scales of 0–12 and 1–5, respectively.

The PSM estimates indicate that the nutritional supplement was associated with a greater likelihood of being currently enrolled in school by 7.8% (95% CI: 0.1%, 15.4%; P < 0.05) and completing 0.8 (95% CI: 0.3, 1.4; P < 0.005) more schooling grades. As with the probit regression models, the results from matching did not show any association between the intervention and test scores.

The findings are robust to alternative matching specifications. With a kernel (Epanechnikov) matching method, the intervention was associated with a 7.9% (95% CI: 2.9%, 12.9%; P < 0.005) rise in current school enrollment and 0.6 (95% CI: 0.3, 0.9; P < 0.001) more schooling grades. Alternatively, after matching each intervention child with 5 nearest neighbors in the control group, the intervention was associated with a 9.3% (95% CI: 3.7%, 14.8%; P < 0.005) current school enrollment rise and a 0.6 (95% CI: 0.3, 1.0; P < 0.005) increase in schooling grades. With a logit model in our propensity score estimation, the intervention was associated with an 8.8% (95% CI: 1.6%, 16.1%, P < 0.05) increase in current school enrollment and 0.7 (95% CI: 0.2, 1.3; P < 0.01) rise in schooling grades. Finally, with the use of the covariate matching technique, the intervention was associated with a 7.2% (95% CI: 1.6%, 12.9%; P < 0.05) rise in current school enrollment and 0.4 (95% CI: 0.1, 0.8; P < 0.05) increase in schooling grade. No significant association with test scores was seen in any of the above estimates.

The basic propensity score estimation model is presented in Table 3, and matching quality tests are presented in Table 4. We found that percentage biases in covariates decreased from a mean of 8% in the unmatched data to 3.7%–5% with PSM. The number of covariates that had statistically significant biases decreased greatly from the unmatched to matched data, whereas the pseudo-R² statistic went down by ≥50% after matching. Furthermore, the likelihood ratio test of joint significance of the covariates was highly statistically significant (P = 0.003) in unmatched data but nonsignificant after matching. These tests confirmed high matching quality and the validity of our method.

TABLE 3.

Estimation of propensity score: probit regression of the likelihood of children belonging to the intervention group in the 2003–2005 follow-up of the Hyderabad Nutrition Trial¹

	School enrollment	Schooling grade completed	Academic performance
Age	0.020 (−0.064, 0.104)	0.022 (−0.064, 0.107)	0.065 (−0.024, 0.162)
Birth order	−0.038 (−0.086, 0.011)	−0.039 (−0.089, 0.011)	−0.045 (−0.102, 0.009)
Male child	−0.005 (−0.162, 0.151)	0.048 (−0.111, 0.207)	−0.157 (−0.342, 0.021)
Scheduled caste/scheduled tribe	0.088 (−0.195, 0.371)	0.043 (−0.248, 0.333)	0.180 (−0.122, 0.495)
Other backward caste	−0.203 (−0.47, 0.064)	−0.264 (−0.538, 0.010)	−0.207 (−0.492, 0.081)
Non-Hindu household	−0.341 (−0.755, 0.073)	−0.390 (−0.807, 0.028)	−0.576 (−1.074, −0.075)
Wealth quintile 2	0.168 (−0.079, 0.416)	0.196 (−0.059, 0.451)	0.117 (−0.163, 0.39)
Wealth quintile 3	0.160 (−0.061, 0.381)	0.134 (−0.093, 0.361)	0.076 (−0.174, 0.326)
Wealth quintile 4	0.175 (−0.066, 0.417)	0.164 (−0.082, 0.409)	0.086 (−0.172, 0.373)
Wealth quintile 5	0.458 (0.196, 0.721)	0.451 (0.184, 0.718)	0.429 (0.14, 0.724)
Father literate	0.002 (−0.218, 0.223)	−0.034 (−0.256, 0.188)	−0.059 (−0.297, 0.173)
Father’s education: primary	−0.261 (−0.507, −0.015)	−0.285 (−0.532, −0.038)	−0.314 (−0.592, −0.046)
Father’s education: secondary and above	−0.259 (−0.538, 0.02)	−0.284 (−0.566, −0.003)	−0.345 (−0.647, −0.051)
Mother literate	0.067 (−0.333, 0.468)	0.039 (−0.362, 0.44)	0.085 (−0.329, 0.495)
Mother’s education: primary and above2	−0.238 (−0.596, 0.121)	−0.255 (−0.615, 0.104)	−0.293 (−0.676, 0.089)
Constant	−0.025 (−1.338, 1.287)	0.022 (−1.312, 1.356)	−0.434 (−1.922, 0.957)
P value for χ2	0.003	0.001	0.000
Sample size, n	1096	1059	904

¹Values are estimated regression coefficients with 95% CIs in parentheses, unless stated otherwise. The number of observations differs across regression models because of some missing values of the outcome variables.

²Because very few mothers have a secondary or higher level of education, we combined them with the category “primary and above.”

TABLE 4.

Percentage of bias in covariates between intervention and control groups in unmatched and matched samples of children in the 2003–2005 follow-up of the Hyderabad Nutrition Trial¹

			Bias after matching
	Bias in unmatched data	P value of bias	School enrollment	Schooling grade completed	Academic performance
Age	4.5	0.46	−2.5	−2.6	6.7
Birth order	−10.2	0.09	7.4	2.6	−0.4
Male child	−1.2	0.84	1.9	3	−0.8
Scheduled caste/scheduled tribe	19.9	0.001	−9.6	−9.2	0.4
Other backward caste	−16.7	0.006	5.5	1.7	−3.4
Non-Hindu household	−12.1	0.042	−4.7	−0.8	2.9
Wealth quintile 2	1.8	0.77	−1.8	0.5	2.5
Wealth quintile 3	0.2	0.98	−2.7	−8.6	−2.7
Wealth quintile 4	−3.2	0.59	8.2	7.1	−6.6
Wealth quintile 5	10.3	0.09	3.8	8.2	6.7
Father literate2	5.6	0.36	13.1	2.3	4.4
Father’s education: primary	−11.4	0.06	−1.9	0	4.4
Father’s education: secondary and above	−8.8	0.15	2.5	6.7	2.2
Mother literate	2.7	0.66	6.5	7.4	2.6
Mother’s education: primary and above3	−10.8	0.07	−2.00	5.4	8.6
Mean bias	8.0		5.0	4.4	3.7
Pseudo-R2	0.023		0.010	0.007	0.006
P-value for χ2	0.003		0.36	0.68	0.90

¹Values are %, unless stated otherwise. Bias is the percentage difference in the mean value of the variable in the intervention sample from that of the control sample. Matching was on propensity scores, using a one-to-one nearest-neighbor match with replacement within a radius of 0.01.

²“Father literate” had a statistically significant bias even after matching (13.1% bias, P < 0.05) in the analysis of enrollment.

³“Mother’s education: primary and above” had a statistically significant bias even after matching (8.6% bias, P < 0.1) in the analysis of academic performance.

In additional analyses, we examined whether nutritional associations were significantly different for girls compared with boys. In the probit regression models of educational outcomes, we included an interaction term between the intervention status and sex of each child (whether female). The estimated coefficient of this interaction term was statistically nonsignificant for the current enrollment and test score regression models, indicating that there was no sex gap in the association between the intervention and these outcomes. However, we found that the intervention was associated with a 0.3 (95% CI: 0.1, 0.6; P < 0.05) greater increase in schooling grade completed by girls compared with boys. In our propensity score matching analysis, we computed the difference-in-difference mean treatment effect on the treated for girls compared with boys (more details are presented in the Supplemental Methods). The results showed that the supplement was associated with a 16.1% (95% CI: 0.9%, 31.2%; P < 0.05) greater rise in current school enrollment of girls. No differential association by sex was seen for the 2 other educational outcomes.

Discussion

More than half—almost 60 million—of the world’s underweight children aged <5 y are Indian (39). The proportion of Indian children who are stunted or underweight declined from 48% and 43%, respectively, in 2005–2006 (29) to 39% and 30% in 2013–2014 (40). However, the prevalences vary considerably across states and districts. In the poorer states of Bihar, Jharkhand, Madhya Pradesh, Orissa, Rajasthan, and Uttar Pradesh, 58.8% and 43.3% of children were still stunted and underweight, respectively, in 2011 (41).

The ICDS has the potential to reduce India’s undernutrition burden, but little is known about its effectiveness. Existing studies of the ICDS use different methods and data from various time periods and reveal positive associations with children’s nutritional status (26, 39, 42–46). However, most of these studies use observational data that may produce biased estimates of causal effects because of systematic differences between beneficiaries and nonbeneficiaries. In addition, there is no evidence, to our knowledge, on associations of ICDS participation with adolescent educational outcomes.

In the broader context of all LMICs, evidence of associations of early childhood nutrition with later-life education is very limited. The most noteworthy studies are from the Institute of Nutrition for Central America and Panama in Guatemala, showing that early life nutritional supplementation was associated with improved schooling grades, reading comprehension, nonverbal skills, wages, and birth weights and heights of the next generation (11, 16, 18, 47). However, it is very unclear that these results from Guatemala in the 1970s generalize to other contexts. Here, we estimated long-term associations of exposure to nutrition in utero and during early life with educational outcomes of Indian adolescents. Using regression and PSM methods, we found that the supplement was associated with consistently higher levels of school enrollment and possibly also with higher schooling grades, although not test scores.

In recent years, India has seen net primary school enrollment rates >90% (48). However, 80 million, or 42% of primary school children, drop out before completing grade 8 (49). Among girls, the dropout rate by grade 10 is as high as 63.5% (50). As our results for 13- to 18-y-olds show, ICDS participation may potentially help reduce these high dropout rates. However, we did not find associations of the intervention with children’s academic performance. Test scores in our study were self-reported and might have suffered from reporting bias. Although unlikely, if the rates of misreporting were different between intervention and control groups, they could distort our findings. Test outcomes are also likely to depend on school quality (51, 52). Sending children for a longer time to inferior quality schools could result in negligible intervention associations with test scores (53). The lack of schooling characteristics data makes such analysis beyond the scope of our study.

There are some limitations of our analysis. First, attrition between the original trial and the 2003–2005 follow-up was nontrivial, albeit lower than that in other relevant studies (26). However, attrition rates in intervention and control groups were not widely different and therefore may not have affected our results. Furthermore, our findings are robust to several econometric model specifications.

Unobserved factors might affect our results in additional ways. After the Hyderabad Nutrition Trial ended, supplementary nutrition was also provided in control villages starting in 1992–1993 (26). Children born in these villages could have received the extra nutrition when they were aged 3–6 y, and this may attenuate associations between the intervention and educational outcomes. However, nutritional interventions are thought to have the strongest effects before age 2–3 y (1, 3).

Access to other public health programs (e.g., immunization and health education) was similar across intervention and control villages. By introducing the nutritional supplement, the ICDS integrated the delivery of such services under a single program. The integration could have increased the uptake of these other components in intervention villages (26), possibly confounding the association between the nutritional supplement and education in our study. Due to lack of data, such effects cannot be separately identified. However, evidence from LMICs with high rates of undernutrition suggests that among early childhood interventions in undernourished populations such as in India, nutrition may be the most important factor to affect future educational outcomes (5, 14, 16, 18). Moreover, the other components of the ICDS, such as immunization and improved hygiene, could reduce infections and therefore increase the effectiveness of the nutritional supplement.

Finally, in the absence of reliable data on actual supplement intakes by participants, we could not conduct a dose-response type analysis whereby associations between duration of exposure to the nutritional supplement and educational outcomes can be tested.

Although we found long-term educational associations with ICDS program participation from a small trial, the population-level program effectiveness depends on its quality and coverage, which vary across children’s age and location (39, 44). The ICDS currently provides 2 types of nutritional supplements. Children aged 3–6 y are given cooked meals or ready-to-eat food for feeding at the Anganwadi centers. Younger children are often given uncooked take-home rations ranging from fortified baby mixtures to ready-to-eat foods (44).

The nutrition provided through cooked meals at Anganwadi centers may be truly supplementary if children or pregnant women continue to receive all or most of their usual food intake at home, as found in Guatemala and the Philippines (54, 55). In comparison, take-home rations may be more extensively reallocated among all family members, thereby increasing energy intakes of women or children only marginally or perhaps not at all.

In 2005–2006, the most recent years for which national data are available (29), coverage of Anganwadi centers ranged from 100% in Tripura to 27% in Meghalaya (national mean of 72%). However, only 26% of children aged <6 y who had neighborhood Anganwadi centers reported receiving nutrition from the centers. Another 2013 report by the Comptroller and Auditor General of India found that >33% of eligible beneficiaries were not receiving supplementary nutrition from the ICDS in many states (56).

In 2013, a multiyear restructuring plan for improving and expanding the ICDS was adopted (23). It included an improved package of services for children aged <3 y and women, enhanced growth tracking, stronger linkages with the public health care system, improvements in infrastructure and staffing at Anganwadi centers, introduction of public-private partnerships, and information campaign activities for greater community participation. Considering the significant educational benefits of the program found in our analysis, along with the health benefits shown by previous studies, the ICDS probably should be restructured and fully funded.