RNA sequencing and other experimental methods that produce large amounts of data are increasingly dominant in molecular biology. However, the noise properties of these techniques have not been fully understood. We assessed the reproducibility of allele-specific expression measurements by conducting replicate sequencing experiments from the same RNA sample. Surprisingly, variation in the estimates of allelic imbalance (AI) between technical replicates was up to 7-fold higher than expected from commonly applied noise models. We show that AI overdispersion varies substantially between replicates and between experimental series, appears to arise during the construction of sequencing libraries, and can be measured by comparing technical replicates. We demonstrate that compensation for AI overdispersion greatly reduces technical variation and enables reliable differential analysis of allele-specific expression across samples and across experiments. Conversely, not taking AI overdispersion into account can lead to a substantial number of false positives in analysis of allelespecific gene expression RNA sequencing (RNA-seq) is a widely used technology for measuring RNA abundance across the whole transcriptome 1 . An especially informative approach to RNA-seq analysis in samples from humans and other diploid organisms is comparison of the activity of the parental alleles. Allele-specific analysis of gene expression can reveal epigenetic gene regulation associated with imprinting 2 , X-chromosome inactivation 3 , and autosomal monoallelic expression 4-8 . The maternal and paternal copies of a gene share the same cell nucleus and therefore are both influenced by the rest of the genome in the same way. Consequently, allelic imbalance (AI) in expression can be highly sensitive to cis-regulatory mechanisms 9,10 . Accordingly, AI analysis has been used to uncover gene regulatory effects in a growing number of studies [11][12][13][14] .Accurate estimation of AI is thus important for quantitative understanding of genetic and epigenetic mechanisms of gene regulation. Efforts to increase AI estimation accuracy have mostly focused on the data analysis 15-20 , based on the assumption that consistency in measuring total RNA abundance translates to accurate measurement of each of the alleles separately. Here, we show that this implicit assumption is incorrect.Based on analysis of newly generated datasets and publicly available data, we reveal a previously neglected, major source of technical variation in AI measurements in RNA-seq experiments. This indicates that analyses based on a single RNA-seq replicate and aiming to identify genes with significant AI can result in an unexpectedly large fraction of false positives. To correct for this issue, we devised a specific measure of AI overdispersion, Quality Correction Constant (QCC), derived from comparison of replicate RNA-seq libraries. QCC can be used to control for this unexpected source of variation, and its use results in more accurate AI estimates. Finally, we outline several use c...