Surrogate variables and the risk of overfitting in neuropsychiatric gene expression analyses
The reproducibility crisis in biomedical research calls for re-evaluation of how common analytic approaches affect results. One such approach, surrogate variable analysis (SVA), is widely used to adjust for unmeasured drivers of variation in transcriptomic differential expression analyses. We systematically quantified the impact of SVA using synthetic data with known ground truth and fourteen brain RNA-seq and microarray datasets spanning Alzheimer's disease, autism spectrum disorder, bipolar disorder, schizophrenia, and Parkinson's disease. In synthetic data, SVA usage frequently led to high false positive rates. This issue was not revealed by standard diagnostics nor resolved by strategies such as maximizing cross-dataset signal correlation. In real data, using SVA reduced cross-cohort reproducibility of differential expression signatures regardless of the number of surrogate variables included. These findings show that SVA risks inducing overfitting at realistic signal levels, advocating for a cautious, case-by-case application, with a preference for modeling known covariates.