Small sample adjustment for inference without assuming orthogonality in a mixed model for repeated measures analysis

Area of Science:

• Biostatistics
• Clinical Trial Methodology
• Longitudinal Data Analysis

Background:

• The mixed model for repeated measures (MMRM) is a common statistical method for longitudinal randomized clinical trials.
• Standard MMRM software often assumes orthogonality between fixed effects and covariance parameters, which may not hold under non-normal error distributions or missing completely at random data.
• This assumption can lead to biased standard errors, particularly in small samples.

Purpose of the Study:

• To address the small-sample bias in standard errors of treatment effects within MMRM analyses.
• To propose novel methods for improving the accuracy of standard error estimation in MMRM.
• To provide accessible software for implementing these improved inference methods.

Main Methods:

• Development of two small-sample adjustment methods for inflating standard errors in MMRM.
• Evaluation of proposed methods through simulation studies.
• Implementation of the proposed methods in an R package for practical use.

Main Results:

• One proposed small-sample adjustment method demonstrated superior performance in reducing underestimation bias of standard errors.
• The simulation results indicate that the recommended method provides empirical conservatism even in general situations.
• The developed R package facilitates easy implementation of these improved inference processes.

Conclusions:

• The proposed small-sample adjustment method is recommended for MMRM analyses when sample sizes are not large and between-group heteroscedasticity is anticipated.
• Accurate standard error estimation is crucial for reliable treatment effect inference in longitudinal studies.
• The study offers a practical solution to improve the statistical rigor of MMRM in challenging scenarios.