Fig. 4

Bayesian regression models for pancreatic cancer risk. To incorporate uncertainty of the copy number assignment from the low-level data, the integer copy number was sampled from the subject-specific posterior probabilities provided by CNPBayes at each iteration of the MCMC. While batch effects on CNV inference were already accounted for in the low and high quality sample collections, an imbalance of the pancreatic cancer cases between these collections warranted a stratified model with an interaction between copy number and data quality and an indicator, z_c, multiplying these coefficients that allowed the slopes to be exactly zero. a Posterior probabilities of association from the stratified model for CNV regions across the genome. For regions where copy number inference was unaffected by data quality and associated with pancreatic cancer risk, regression coefficients for the low and high quality collections were positively correlated and the posterior mean of z_c (upper right corner) increased in the more powerful unstratified analysis using all 7598 samples (b). By contrast, negatively correlated coefficients indicated an effect of data quality on CNV inference confirmed by visual inspection and the appropriate follow-up analysis and estimated probability of association was limited to the high quality sample collection (c)