Robust small-sample inference for fixed effects in general Gaussian linear models.

Area of Science:

• Biostatistics
• Statistical Modeling
• Longitudinal Data Analysis

Background:

• Empirical covariance estimators are asymptotically consistent but can exhibit significant bias in small samples.
• Existing bias-correction methods may not adequately address small-sample issues in general Gaussian linear models.

Purpose of the Study:

• To propose a novel small sample correction for the empirical covariance estimator in general Gaussian linear models.
• To derive inference for fixed effects using the corrected covariance matrix.
• To evaluate the performance of the proposed method against existing bias-correction techniques.

Main Methods:

• Development of a new bias-correction formula for the empirical covariance estimator.
• Application of the corrected estimator to derive fixed-effects inference.
• Illustrative examples using a two-way analysis of variance (ANOVA) model with repeated measures and a four-period crossover design.

Main Results:

• The proposed small sample correction effectively reduces bias in covariance estimation for small sample sizes.
• Inference for fixed effects based on the corrected covariance matrix is established.
• Simulation studies indicate that the proposed method outperforms existing bias-correction approaches (Mancl and DeRouen, Kauermann and Carroll, Fay and Graubard) in balanced designs.

Conclusions:

• The proposed small sample corrected covariance estimator offers improved accuracy for Gaussian linear models with limited data.
• This method provides a valuable tool for robust statistical inference in complex study designs like repeated measures and crossover trials.
• The findings suggest a more reliable approach to analyzing data from studies with small sample sizes.