Statistical analysis

We used an online analysis tool - SNP Effect Concordance Analysis (SECA; http://neurogenetics.qimrberghofer.edu.au/SECA/) - to remove SNPs in linkage disequilibrium (LD) (r² > 0.1) within 10 Mb and kept a set of independent SNPs with the lowest P values in overall meta-analysis. Among these independent SNPs, we further selected those that reached nominal significance level of P < 0.05 in both GWAS datasets with consistent direction of association as index SNPs. We performed eQTL analysis using the Genotype-Tissue Expression (GTEx) Portal (http://www.gtexportal.org/home/). Genotype data from 1000G Phasel v3 CEU (b37 rsIDs) was used for LD estimation. To evaluate interactions between sun exposure and genotypes, we modeled sun exposure level as a continuous variable using the median value among controls for each tertile in the Harvard cohort, which allowed us to assess the statistical significance of interaction by likelihood ratio tests. We also constructed a multivariable confounder score to summarize pigment traits in the Harvard cohort.²⁵ Briefly, we applied the logistic regression coefficients from a multivariable model for skin cancer risk, including age, gender, natural hair color, times of blistered sunburn and tendency to sunburn during adolescence, to each individual’s values for the latter three of these variables and summed the values to compute a pigment score in the logit scale. We used the median value for this score among controls to identify participants with low, intermediate, and high pigmentation. All statistical tests were two-sided.