An evaluation of approaches for rare variant association analyses of binary traits in related samples

Recognizing that family data provide unique advantage of identifying rare risk variants in genetic association studies, many cohorts with related samples have gone through whole genome sequencing in large initiatives such as the NHLBI Trans-Omics for Precision Medicine (TOPMed) program. Analyzing rare variants poses challenges for binary traits in that some genotype categories may have few or no observed events, causing bias and inflation in commonly used methods. Several methods have recently been proposed to better handle rare variants while accounting for family relationship, but their performances have not been thoroughly evaluated together. Here we compare several existing approaches including SAIGE but not limited to related samples using simulations based on the Framingham Heart Study samples and genotype data from Illumina HumanExome BeadChip where rare variants are the majority. We found that logistic regression with likelihood ratio test applied to related samples was the only approach that did not have inflated type I error rates in both single variant test (SVT) and gene-based tests, followed by Firth logistic regression that had inflation in its direction insensitive gene-based test at prevalence 0.01 only, applied to either related or unrelated samples, though theoretically logistic regression and Firth logistic regression do not account for relatedness in samples. SAIGE had inflation in SVT at prevalence 0.1 or lower and the inflation was eliminated with a minor allele count filter of 5. As for power, there was no approach that outperformed others consistently among all single variant tests and gene-based tests.