Online Supplement to Polygenic Influence on Educational Attainment. Genotyping was conducted with the Illumina HumanOmni1-Quad v1 platform using

Transcription

1 Online Supplement to Polygenic Influence on Educational Attainment Construction of Polygenic Score for Educational Attainment Genotyping was conducted with the Illumina HumanOmni1-Quad v1 platform using DNA extracted (via Oragene saliva collection) from individuals at Wave 4. Complete details of the QC process that resulted in the data used here are available in McQueen et al. (2014). We removed 18,665 SNPs from the original panel of genetic data (based on a missingness threshold of 5% computed in a sample of individuals who contained information on at least 90% of all available SNPs) to arrive at a genetic database consisting of 940,862 SNPs. In the original QC analysis, 74 genetic samples (some of which may have been duplicates included as part of the QA process) were dropped due to missingness concerns. SNPs in the Add Health Sibling Pairs genetic database were matched to SNPs with reported results in the educational attainment GWAS (Rietveld et al., 2013). 1 Over 2/3 of the SNPs in the Add Health genetic database were included in the GWAS results (642,627 SNPs). For each of these SNPs, a loading was calculated as the number of education associated alleles multiplied by the effect-size estimated in the original GWAS. SNPs with relatively large p-values will have small effects (and thus be down weighted in creating the composite), so we do not impose a p-value threshold. Research has suggested that accounting for linkage disequilibrium (LD) structure can improve the predictive performance of polygenic scores (Vilhjalmsson et al., 2015). To test the sensitivity of the score, we computed a secondary score based on a randomly chosen sample of unrelated EA respondents (N=507). The matched set of SNPs was pruned to account for linkage disequilibrium using the clumping procedure (which considers the level of association 1 Results are publicly available, 1

2 between the SNP and the phenotype, not simply LD) in the second-generation PLINK software (Chang et al., 2014). Clumping takes place in two steps. 2 The first pass is done in fairly narrow windows (250kb) for all SNPs (the p-value significance thresholds for both index and secondary SNPs is set to 1) with a liberal LD threshold (r 2 =0.5). In a second pass, SNPs remaining after the first prune are again pruned in broader windows (5000kb) but with a more conservative LD threshold (r 2 =0.2). SNPs are then weighted based on effect sizes reported in the GWAS as discussed above. Amongst the EA respondents, the clumped score was correlated with the original score at Again amongst the EA respondents, correlation with educational attainment was slightly higher for the un-clumped score, r=0.18 (see Table 2 of main text), as compared to the correlation of educational attainment with the clumped score (r=0.15). Given the fact that the clumped score did not improve predictive performance and since clumping depends upon LD patterns that may vary across samples, we report results for the un-clumped score in the main text. Construction of Neighborhood Disadvantage Index We constructed a measure of neighborhood disadvantage using data from an individual s census block group at the baseline interview. 3 Construction of this variable was performed as follows. In a first step, we identified those contextual variables associated with educational attainment in the full Add Health sample (p<0.05 when regressed on educational attainment). We then conducted factor analysis (using the algorithm of Stacklies et al., 2007) of this subset of 2 We used thresholds suggested by Sarah Medland. Please see 3 In particular, we used the 29 variables described here: 2

3 contextual variables to generate a neighborhood disadvantage factor score based on the first principal component. Loadings for the first principal component, which explained 19.2% of the variance, can be found in Table S1. Table 1 shows that those amongst both EA and AA respondents, our analytic sample contains individuals from more disadvantaged neighborhoods than those in the full subsample of the AH cohort. Calculation of Power for sibling-based analyses We conducted a power analysis of our ability to detect an effect in the sibling analyses in the following manner. We first restricted our sample to those EA respondents in sibling pairs (N=772). We then residualized educational attainment on birthyear so as to simplify the subsequent analyses. For residualized attainment, we estimated β W from model 3 to be coefficient of We then standardized this using the SD of the residualized attainment and obtained Finally, we simulated data accounting for the clustering of both scores and attainment within families. For individual i in family j, we generate data via the following two equations: y ij = b score ij + ε ij score ij = μ j + ε ij where the distributions of μ j, ε ij, and ε ij are based on the empirical data. We generated 500 datasets for different values of b. Results are shown in Figure S2. The observed coefficient of 0.28 corresponds to a power of approximately

4 References Chang, C. C., Chow, C. C., Tellier, L. C., Vattikuti, S., Purcell, S. M., & Lee, J. J. (2014). Second-generation PLINK: rising to the challenge of larger and richer datasets. arxiv preprint arxiv: Domingue, B. W., Belsky, D. W., Harris, K. M., Smolen, A., McQueen, M. B., & Boardman, J. D. (2014). Polygenic risk predicts obesity in both white and black young adults. PloS one, 9(7), e McQueen, M. B., Boardman, J. D., Domingue, B. W., Smolen, A., Tabor, J., Killeya-Jones, L.,... & Harris, K. M. (2014). The National Longitudinal Study of Adolescent to Adult Health (Add Health) Sibling Pairs Genome-Wide Data. Behavior genetics, Rietveld, C. A., Medland, S. E., Derringer, J., Yang, J., Esko, T., Martin, N. W.,...& McMahon, G. (2013). GWAS of 126,559 individuals identifies genetic variants associated with educational attainment. Science, 340(6139), Stacklies, W., Redestig, H., Scholz, M., Walther, D., & Selbig, J. (2007). pcamethods a bioconductor package providing PCA methods for incomplete data. Bioinformatics, 23(9), Vilhjalmsson, B., Yang, J., Finucane, H. K., Gusev, A., Lindstrom, S., Ripke, S.,... & Schizophrenia Working Group of the Psychiatric Genomics Consortium. (2015). Modeling Linkage Disequilibrium Increases Accuracy of Polygenic Risk Scores. biorxiv,

5 Table S1. Factor loadings for contextual variables (taken from respondent s census block group) used to create neighborhood disadvantage. In a first stage, contextual variables unassociated with educational attainment were screened out. The second block shows variables which were transformed from their original coding. All variables were standardized to have mean of 0 and SD of 1. Variable Description Loading BST90P04 Proportion hispanic 0.11 BST90P06 Median age BST90P07 Dispersion in age distribution BST90P10 Propotion population < 5 years old 0.12 BST90P12 Dispersion in migration status 0.02 BST90P15 Median household income BST90P16 Dispersion in household income 0.08 BST90P17 Median family income BST90P18 Dispersion in family income 0.18 BST90P19 Proportion persons with income < poverty level 0.37 BST90P20 Modal educational attainment BST90P21 Dispersion in educational attainment

6 BST90P22 Proportion females in labor force BST90P23 Unemployment rate 0.29 BST90P26 Tenure of occupied housing units BST90P27 Proportion occupied housing units moved into between 1985 and BST90P28 Median value of housing units BST90P29 Dispersion in value of housing units BST90P02 Modal race = black 0.22 BST90P02 Modal race = other 0.00 BST90P08 Modal marital status = married BST90P08 Modal marital status = divorced 0.05 BST90P13 Modal housedhold type = other 0.20 BST90P13 Modal household type = non-family 0.09 BST90P24 Modal occupation type = technical BST90P24 Modal occupation type = service 0.15 BST90P24 Modal occupation type = production 0.04 BST90P24 Modal occupation type = laborers

7 Figure S1. Top row: Density plots of polygenic scores for EA and AA respondents. Middle row: Histograms of birth years and educational attainment at wave 4 as well as a scatterplot (r=-0.08) comparing the two for the genotyped EA subsample of respondents. Bottom row: Histograms of birth years and educational attainment at wave 4 as well as a scatterplot (r=0.03) comparing the two for the genotyped AA subsample of respondents. 7

8 Figure S2. Power curve for the sibling-based analyses. The vertical line is the observed estimate (b=0.28) in a sibling-based analysis of residualized educational attainment. 8