|
Abstract:
|
In repeated measurement analyses, variance-covariance structure is considered having potential impact on modeling main interests, estimating/testing fixed effects. Choosing appropriately is no easy task. Relying on information criteria such as AIC/AICC/BIC is a common practice but with known limitations, such as overfitting/underfitting. Type I error inflation is often observed. Or a pre-assumed general variance-covariance structure is used, such as un-structured, without consulting data to avoid selection. Such over-generality could lead to significant loss of power, especially for studies with relatively small sample sizes. Simulations were conducted for typical phase I settings, impacts of above two common practices were explored with attention in balancing type I error inflation and loss of power. Benefit of heterogeneity among treatments if pre-knowledge suggests was investigated.
|
