Patterns of growth in childhood in relation to adult schooling attainment and IQ in 6 birth cohorts in low and middle-income countries: evidence from COHORTS

Background: Growth faltering has been associated with poor intellectual performance. The relative strengths of associations between growth in early and in later childhood remain underexplored. We examined the association between growth in childhood and adolescence and adult human capital in five low- or middle-income countries (LMICs). Methods: We analyzed data from six prospective birth cohorts of five LMICs (Brazil, Guatemala, India, the Philippines, and South Africa). We assessed the associations of measures of height and relative weight at four ages (birth, at around age 2 years, mid-childhood (MC), adulthood), with two dimension of adult human capital (schooling attainment and IQ). Findings: In site- and sex-pooled analyses, size at birth and linear growth from birth to around 2 years of age were positively associated with schooling attainment and adult IQ. Linear growth from age 2 years to MC and from MC to adulthood was not associated with higher school attainment or IQ. Change in relative weight in early childhood was not associated with either outcome. Relative weight in MC and in adulthood were inversely associated with schooling attainment but were not associated with adult IQ. Interpretation: Linear growth in the first 1,000 days is a predictor of schooling attainment and IQ in adulthood in LMICs. Linear growth in later periods was not associated with either of these outcomes. Changes in relative weight had inconsistent association with schooling and IQ in adulthood.


0
All cohort-specific analyses were performed for males and females separately. We stratified the 1 9 9 analyses by study site and sex given observed heterogeneity among sites and previous literature 2 0 0 that supports sex differences in IQ scores. 31 Sex-combined estimates were generated by pooling 2 0 1 the sex-specific estimates using weighted random effects meta-analysis, with the weight each sex 2 0 2 received being proportional to sample size.

0 3
A doubly-robust strategy was used for covariate adjustment, where adjustment was performed 2 0 4 via multivariable regression after inverse probability of treatment weighting (IPTW) using the 2 0 5 "ipwpoint" function from the "ipw" package in R, 32 applying linear regression to model the 2 0 6 relationship between the exposure variable and the covariates. This regression was specified to 2 0 7 include a "main effect" term for all covariates, as well as all pairwise product terms between influence the results, the left tail of the weights was truncated at the 0.5th percentile, and the 2 1 0 right tail at the 99.5th percentile.

1 1
We defined four models a priori with progressive adjustment of potential confounders. We used 2 1 2 listwise deletion in all models. In model one (minimal adjustment), we adjusted for sex (for two, we controlled for the same variables as in model one but excluded cases that could not be 2 1 5 included in subsequent models because of missing values. We found similar point estimates 2 1 6 between models one and two, suggesting that missing covariate data was not a major factor in 2 1 7 our results. In model three, we adjusted for the covariates in model one, plus early-life 2 1 8 socioeconomic quintiles, maternal schooling, maternal age, maternal height, birth order and, for 2 1 9 the Brazil 1982 and 1993 cohorts, skin color. Comparing models two and three allowed us to 2 2 0 assess changes due to confounding by the selected covariates. In model four (further adjustment), we controlled for covariates in model three plus paternal schooling. Model four is our preferred 2 2 2 representation.

3
We used random effects meta-analysis to pool the sex-and cohort-specific results. Pooled sex-2 2 4 combined estimates were generated by pooling the corresponding pooled sex-specific estimates. The variation between cohorts was estimated using I² statistic and Cochran's Q test, and random 2 2 6 effects meta-regression was used to test for the effect modification by sex. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted June 18, 2020. . The funders of this study did not have any role in study design, collection and analysis of data, 2 3 1 description and interpretation of results, and writing of this manuscript. The corresponding 2 3 2 author had full access to all the data in the study and had final responsibility for the decision to 2 3 3 submit for publication. Data from 9503 participants with complete data for at least one of the outcomes and size at 2 2 3 8 years of age were analyzed. Table 1 shows selected characteristics of the participants. Birth Indian participants had the lowest height-for-age and weight-for-age z-scores in early and mid-2 4 1 childhood and were shortest as adults. Schooling for both parents was lowest in Guatemala.

4 2
School attainment was higher in women across all cohorts when compared to men, except in  The site-and sex-pooled associations between growth and school attainment are presented in    Conditional height in adulthood was not associated with schooling attainment in men (p=0.48) 2 5 3 but it was inversely associated in women (p=0.013). Conditional relative weight at around 2 2 5 4 years was not associated with schooling attainment in fully adjusted models. Conditional relative were generally consistent between males and females and across the six cohorts, although there 2 5 7 was heterogeneity in the size of the estimates (Table 3). Sex differences were observed in the 2 5 8 estimates for conditional height in adulthood, and for the estimates for conditional relative 2 5 9 weight in MC and in adulthood. In models that examined the binary categorization of schooling 2 6 0 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

5 1
Among the study's limitations, we note some inconsistencies in the ages of exposures and 3 5 2 outcomes across the cohorts. Height was measured at 2 years of age in all sites except Brazil in most of the cohorts, but 8.5 years in the Philippines. There were also differences in the ages in 3 5 5 which the schooling and IQ outcomes were obtained, and some differences in the instruments 3 5 6 used to measure IQ across sites. Finally, residual and unmeasured confounding should be 3 5 7 considered given the observational nature of our study.

5 8
This study has several strengths. We analyzed six well-characterized population-based birth 3 5 9 cohorts in five LMIC with follow-up periods that ranged from 18 to 46 years. Each study site had 3 6 0 trained staff who followed standard methodologies to collect the anthropometric and intelligence, and we standardized the distributions within each cohort and by sex to be able to 3 6 5 compare outcomes. We adjusted our models for a range of early life social factors. Additionally,  Table 1). We saw evidence of heterogeneity among the cohorts in the magnitude 3 6 9 of the estimates but not in the direction of the associations. Thus, we were able to obtain single 3 7 0 pooled estimates combining cohort-specific results. These associations might be generalized to 3 7 1 other LMICs, but should be interpreted with caution given the differences between pooled and 1 6 years, during childhood, adolescence or adulthood, were associated with adult human capital    Policy Report 1996; X(5): 32. . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted June 18, 2020.

2 0
Sample sizes refer to participants with non-missing data for at least one outcome and non-missing data for either HAZ or WAZ at 2 years of age. Adult age was 5 2 1 estimated based on date of IQ measurement, except India (age at latest wave of data collection). is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted June 18, 2020.  sexes and in Guatemala analysis, we controlled for year at birth and intervention group variables. In fully adjusted analyses, we also controlled for maternal 5 2 7 factors (height, age at childbirth, schooling), paternal schooling, birth order, and income/wealth quintiles. Additionally, we controlled for maternal skin color in 5 2 8 both Brazil cohorts. IQ = intelligence quotient. *p-values for associations < 0.05. is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted June 18, 2020.   site-specific were calculated using birth weight as anchor. In adjusted analyses, we controlled for maternal factors (height, age at childbirth, schooling), birth 5 4 1 order, and income/wealth quintiles. Additionally, we controlled for year at birth and intervention group variables (in Guatemala analysis), and for maternal skin color (in both Brazil cohorts). These models exclude cases with missing values. The I² statistic and Cochran's Q test were used to quantify between-cohort is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted June 18, 2020.