multtest

multtest implements resampling-based multiple testing procedures to control Type I error rates in high-dimensional genomic studies such as microarray and other high-throughput assays.

Key Features:

• Resampling frameworks: Supports nonparametric bootstrap and permutation resampling frameworks and empirical Bayes–based approaches.
• Error-rate control: Implements single-step and stepwise algorithms to control family-wise error rate (FWER), generalized FWER (gFWER), tail probability of the proportion of false positives (TPPFP), and false discovery rate (FDR).
• Test statistics and models: Constructs tests using t- and F-statistics derived from linear models, regression coefficients, correlation parameters, and Cox proportional hazards models, accommodating censored survival data.
• Bootstrap null distributions: Offers centered, centered-and-scaled, and quantile-transformed bootstrap null distributions.
• Multivariate normal approximation: Provides a computationally efficient null distribution based on a multivariate normal approximation with covariance derived from the vector influence function.
• Outputs: Produces adjusted p-values, confidence regions, and test-statistic cutoffs tailored to the selected error-rate control procedure.

Scientific Applications:

• Large-scale genomic studies: Applies resampling-based multiple testing to microarray and other high-throughput assays for large-scale hypothesis testing.
• Survival analysis: Identifies genes associated with censored survival outcomes using Cox proportional hazards models and resampling-based inference.
• Multivariate clinical associations: Tests associations between gene expression and multivariate clinical covariates via linear models, regression coefficients, and correlation parameters.
• Case study: Applied to the ALL study of Chiaretti et al. to identify genes whose expression patterns associate with complex phenotypes.

Methodology:

Performs nonparametric bootstrap and permutation resampling and empirical Bayes approaches, uses centered/centered-and-scaled/quantile-transformed bootstrap nulls, computes t- and F-statistics from linear models, regression coefficients, correlation parameters, and Cox proportional hazards models, and optionally uses a multivariate normal approximation with covariance from the vector influence function.