614 – Mixed- and Random-Effects Modeling
Comparing Methods of Small Sample Inference for Linear Mixed Models
Guixian Lin
SAS Institute
Min Zhu
SAS Institute
Linear mixed models have become a popular framework for analyzing data that have complicated designs and structured covariances. No exact test is available for fixed effects except for balanced data and special covariance structures. Test statistics based on asymptotic distribution of parameter estimates are useful for large samples. However, these tests can be unreliable in applications with small sample sizes. Various methods have been proposed to better approximate small sample distributions of these test statistics. This paper compares the performance of these methods through simulation studies that vary in experiment design, covariance structure, sample size, and so on. It also investigates robustness to misspecification of covariance structures and nonnormal random effects.
