Adjustment Procedure to Permutation Tests in Epigenomic Differential Analysis
MetadataShow full item record
In the analysis of genomic data, t-statistics are widely used to detect differential signals between different groups of samples. In many studies, each group has only a small number of replicate samples, making the variance estimation unstable. Small sample variances due to chance can create large t-statistics for genes or genomic loci that are not differential. In order to mitigate this problem, shrinkage estimators are now widely used for variance estimation. One example is moderated t-statistics. For statistical inference, null distributions need to be constructed for test-statistics. Permutation is a natural option to construct null distributions when they cannot be derived using a parametric model due to violations of parametric assumptions. When variance shrinkage estimators are involved, naive permutation can be misleading. This is because for a differential gene or locus, permuting measurements between two groups will inflate the variance estimate which in turn will influence the variance shrinkage estimator. This thesis investigates this issue and proposes a solution to this problem by permuting residuals. This approach is applied and evaluated in genomic applications that involve comparisons of one or multiple data types between two biological conditions.