Non-random sampling leads to biased estimates of transcriptome association

Integration of independent data resources across -omics platforms offers transformative opportunity for novel clinical and biological discoveries. However, application of emerging analytic methods in the context of selection bias represents a noteworthy and pervasive challenge. We hypothesize that combining differentially selected samples for integrated transcriptome analysis will lead to bias in the estimated association between predicted expression and the trait. Our results are based on in silico investigations and a case example focused on body mass index across four well-described cohorts apparently derived from markedly different populations. Our findings suggest that integrative analysis can lead to substantial relative bias in the estimate of association between predicted expression and the trait. The average estimate of association ranged from 51.3% less than to 96.7% greater than the true value for the biased sampling scenarios considered, while the average error was − 2.7% for the unbiased scenario. The corresponding 95% confidence interval coverage rate ranged from 46.4% to 69.5% under biased sampling, and was equal to 75% for the unbiased scenario. Inverse probability weighting with observed and estimated weights is applied as one corrective measure and appears to reduce the bias and improve coverage. These results highlight a critical need to address selection bias in integrative analysis and to use caution in interpreting findings in the presence of different sampling mechanisms between groups.