GWAS中极端不平衡数据统计分析方法

Extremely imbalanced data here refers to datasets where the values of independent or dependent variables exhibit severe imbalances in proportions, such as extremely imbalanced case-control ratios, very low disease incidence rates, heavily censored survival data, and low-frequency or rare variants. In such scenarios, test statistics in classical statistical methods, such as logistic regression and Cox proportional hazards models, may deviate from normality assumptions, leading to difficulties in controlling type I errors. With the increasing availability and exploration of resources from large-scale population cohorts in whole-genome association studies, there is a growing demand for efficient and accurate statistical approaches to handle extremely imbalanced data in independent and non-independent samples. To address this need, this paper provides a systematic methodological overview. Firstly, it derives test statistics from classical statistical methods. Secondly, it elucidates the impact of extremely imbalanced data on the distribution of test statistics. Thirdly, it introduces two widely used methods for correcting statistics in genome-wide association studies: Firth correction and saddlepoint approximation methods. Finally, it briefly introduces commonly used software for extremely imbalanced genomic data. This paper provides theoretical references and application recommendations for the statistical analysis of extremely imbalanced data.