3039 — Modeling Strategies for Handling Baseline Values in Longitudinal Randomized Controlled Trials
Coffman CJ, and Edelman D, Durham VA Medical Center, HSR&D; Fredrickson SK, Hunter Holmes McGuire VA Medical Center; Jeffreys AA, and Woolson RF, Durham VA Medical Center, HSR&D;
Often in longitudinal randomized controlled trials (RCT) each patient is measured at the same set of occasions, and the number of follow-up measurement occasions is few, e.g. 2 to 4. In these trials baseline outcomes are generally measured prior to randomization and can be incorporated in the analysis in a variety of ways that directly impact the results. In this study, we fit six different plausible models to evaluate the treatment effect of a Group Medical Clinic intervention on LDL cholesterol from the VA Group Visits RCT on diabetic patients. We highlight the differences between the models and clarify the research questions addressed by each model. In addition, we examine the impact of missing data on each of the models.
The six models we fit can be divided into two classes: conditional and unconditional. A conditional model or Analysis of Covariance (ANCOVA) approach yields estimates of treatment differences over time, conditional on the observed baseline values. An unconditional model, i.e., a constrained longitudinal model (cLDA) or a longitudinal model (LDA) yields estimates of treatment differences unconditionally. Both cLDA and LDA models use all available data for a subject and will yield unbiased estimates under the assumption that missing outcomes are ignorable. In contrast, for ANCOVA models, when only the baseline value is obtained, the subject is deleted from the analysis.
Fasting lipid profiles on 239 Veterans were measured at baseline, midpoint and end of study. LDL-C was missing for 15, 37, and 35 Veterans at baseline, midpoint, and end of study, respectively. Treatment effect differences ranged from 7.8 to 11.8 mg/dl between the intervention and usual care groups using all available data (n = 239) and ranged from 11.4 to 13.1 mg/dl when analyzing completers only (n = 182). Standard error estimates also varied.
Commonly-used methods of longitudinal data analysis can lead to different interpretations of treatment effects on LDL-C that reflect fundamental assumption differences between conditional and unconditional models, and how missing data are handled.
To guide model selection for analysis of RCT longitudinal trials it is critical to define research questions precisely so that the appropriate and most robust analysis is performed.