Both sample and CNV-specific QC was conducted by examining distributions of QC metrics informed by comparable published CNV analyses, e.g.,22,24,25. Since distributions were similar for calls from PennCNV and iPattern, the same QC thresholds were used for both algorithms to maximize comparability (see Supplementary Methods, available online). QC-filtered PennCNV and iPattern callsets were merged at the sample level using CNVision (http://futo.cs.yale.edu/mw/index.php/CNVision). Only calls with >50% overlap based on the union of the CNV region were included in the analysis. Overlap percentages were higher for the Hap1M (86%–98%) compared to the Hap610 chip (59%–82%; Table S1, available online). Due to the presence of a batch effect within the Hap610 samples, analyses were restricted to large (>500kb) events, the size at which batch effects were no longer observed (Figure S1, Table S2, available online). All CNVs were also filtered for rare events (<1% frequency in the Database of Genomic Variants).
