Abstract:
|
The goal of the current study is to compare approaches that allow for flexible missing data mechanisms when analyzing treatment effects and moderators from pre-post data in RCTs. In general, maximum likelihood and multiple imputation methods are considered the state-of-the-art for analysis when data are missing (Schafer, 2002, Psychological Methods). Nevertheless, little research appears to be available on these methods when analyzing treatment effects and moderators from pre-post data in RCTs. We investigate an extension of the mixed model approach (see, e.g., Winkens et al, 2007, Contemporary Clinical Trials) for maximum likelihood analysis and the multiple imputation approach to analysis of covariance and its non-parallel regression line extension for the analysis of treatment effects and moderators. These models are introduced and compared to determine their underlying operating characteristics within the practice of the analysis of RCTs from pre-post data, focusing on (a) model specification for the analysis of treatment effects and moderators, (b) the effect of random covariates on their statistical properties, and (c) the small-sample properties of these methods.
|