Comparing transcript levels between healthy and diseased individuals allows the identification of differentially expressed genes, which may be causes, consequences or mere correlates of the disease under scrutiny. We propose a method to decompose the observational correlation between gene expression and phenotypes driven by confounders, forward- and reverse causal effects. The bi-directional causal effects between gene expression and complex traits are obtained by Mendelian Randomization integrating summary-level data from GWAS and whole-blood eQTLs. Applying this approach to complex traits reveals that forward effects have negligible contribution. For example, BMI- and triglycerides-gene expression correlation coefficients robustly correlate with trait-to-expression causal effects (r _BMI = 0.11, P _BMI = 2.0 × 10 ^-51 and r _TG = 0.13, P _TG = 1.1 × 10 ^-68 ), but not detectably with expression-to-trait effects. Our results demonstrate that studies comparing the transcriptome of diseased and healthy subjects are more prone to reveal disease-induced gene expression changes rather than disease causing ones.