We used a Fisher’s exact test framework for our variant-level and gene-level analyses based on previous success with this approach16–23. Limitations of the Fisher’s test compared with regression-based approaches12,34,44,45 include an inability to adjust for covariates. On a subset of traits selected for comparisons, we observed that the Phred scores for significant variants from the Fisher’s exact test, SAIGE SPA and REGENIE 2.0.2 were nearly perfectly correlated (Pearson’s r = 0.99). The Fisher’s exact test generated more conservative statistics for rare variants and was associated with increased computational efficiency. Use of the Fisher’s exact test requires extremely careful quality control, case–control harmonization and ancestry pruning. In the absence of these measures, it is crucial to correct for such confounders via a regression-based approach. Future work should focus on in-depth benchmarking for these different methods. Regardless of the approach used, it is essential to define an appropriate study-wide significance threshold, which we addressed using n-of-1 permutation and an empirical null distribution using a synonymous negative control model.
