|
Abstract:
|
In randomized clinical trials, the outcome variable is usually measured prior to randomization (baseline). Eligibility for study participation sometimes requires this baseline value to exceed a threshold. The outcome variable is measured at prespecified intervals over the course of the study, so that each subject has a panel of measurements over time. A common approach is to model within-subject correlation using random effects, and to assess treatment differences by focusing on a model for treatment by time interaction. When the degree of response depends upon the level of pathology at baseline, primary interest generally focuses on treatment differences net of baseline effects. We examine common approaches for baseline adjustment, and demonstrate how popular methods can sacrifice power, validity, and/or numerical stability. We also examine alternative approaches.
|
