Abstract
We present the Causal Pivot (CP) as a structural causal model (SCM) for analyzing genetic heterogeneity in complex diseases. The CP leverages an established causal factor or factors to detect the contribution of additional suspected causes. Specifically, polygenic risk scores (PRSs) serve as known causes, while rare variants (RVs) or RV ensembles are evaluated as candidate causes. The CP incorporates outcome-induced association by conditioning on disease status. We derive a conditional maximum-likelihood procedure for binary and quantitative traits and develop the Causal Pivot likelihood ratio test (CP-LRT) to detect causal signals. Through simulations, we demonstrate the CP-LRT’s robust power and superior error control compared to alternatives. We apply the CP-LRT to UK Biobank (UKB) data, analyzing three exemplar diseases: hypercholesterolemia (HC, low-density lipoprotein cholesterol ≥4.9 mmol/L; nc = 24,656), breast cancer (BC, ICD-10 C50; nc = 12,479), and Parkinson disease (PD, ICD-10 G20; nc = 2,940). For PRS, we utilize UKB-derived values, and for RVs, we analyze ClinVar pathogenic/likely pathogenic variants and loss-of-function mutations in disease-relevant genes: LDLR for HC, BRCA1 for BC, and GBA1 for PD. Significant CP-LRT signals were detected for all three diseases. Cross-disease and synonymous variant analyses serve as controls. We further develop ancestry adjustment using matching and inverse probability weighting as well as regression and doubly robust methods; we extend this to examine oligogenic burden in the lysosomal storage pathway in PD. The CP reveals an approach to address heterogeneity and is an extensible method for inference and discovery in complex disease genetics.