We applied a mixed-model approach to test for heterogeneity in eQTL signal strength. We assume that expression data from the same individual, and from the same exon within a gene, may be correlated. To test for heterogeneity among probe sets in a gene, we fitted a no-heterogeneity model containing a SNP allele dosage main effect and two random effects terms indexing each individual in the data set and each exon in the gene. We then fitted a heterogeneity model containing the same terms plus a set of fixed-effect SNP dosage × exon ID interaction terms. Both models were fitted via restricted maximum likelihood using the lmer() function in the lme4 R package. A likelihood ratio test was used to determine significance. To ensure that significant heterogeneity was not being driven by the absence of expression in all but one probe set, the test was only conducted in transcripts with at least three probe sets that either had a significant eQTL signal or were considered to be expressed above background (significant DABG).
