Exact two-sample inference with missing data.

Area of Science:

• Biostatistics
• Statistical Inference
• Data Analysis

Background:

• Missing data in longitudinal studies can lead to biased inferences if dropout is related to baseline covariates.
• Standard statistical methods may fail when follow-up data is incomplete and censoring is covariate-dependent.

Purpose of the Study:

• To develop an exact, nonparametric inference method for U-statistics with covariate-dependent dropouts.
• To address potential bias arising from correlated follow-up measurements and baseline covariates in incomplete datasets.

Main Methods:

• The proposed method uses a class of modified U-statistics with exact inference.
• It involves weighting permutations by estimated retention probabilities to account for the missing data mechanism.
• Nonparametric approach requires no distributional assumptions for outcomes or missingness patterns.
• Monte Carlo approximation via Gibbs sampling is employed for computational efficiency.

Main Results:

• The Gibbs sampler-based Monte Carlo approximation is demonstrated to be fast and accurate through simulations.
• The method provides exact inference even when asymptotic procedures are unsuitable.
• Illustrative examples using small datasets are presented.

Conclusions:

• The developed nonparametric method offers a robust solution for statistical inference in the presence of covariate-dependent dropouts.
• Accurate analysis is achievable without assuming specific data distributions.
• The technique is particularly valuable for small datasets where traditional methods may be inappropriate.