Accounting for Baseline Covariates and Missing Data in Regulatory Trials with Longitudinal Designs

• Activity Number: 357 - SPEED: Biopharmaceutical Statistics
• Type: Contributed
• Date/Time: Tuesday, August 1, 2017 : 10:30 AM to 11:15 AM
• Sponsor: Biopharmaceutical Section
• Author(s): Elizabeth Colantuoni* and Jon Steingrimsson and Aidan McDermott and Michael Rosenblum
• Companies: Johns Hopkins University and Johns Hopkins University and Johns Hopkins University and Johns Hopkins University
• Keywords: clinical trial ; longitudinal design ; baseline covariates ; patient drop-out ; mixed model for repeated measurement ; targeted minimum loss estimators

Abstract

Consider a two arm regulatory trial where the primary outcome is measured at baseline and several fixed follow-up times. The primary endpoint is the change in the primary outcome from baseline to the final follow-up and the average treatment effect is the difference in the mean change comparing the treatment and control arm. Assume that a set of potentially prognostic baseline variables are collected and patient drop-out is expected. To estimate the average treatment effect, the mixed model for repeated measurement (MMRM) is the standard statistical approach. However, novel targeted minimum loss estimators (TMLE) proposed by Van der Laan and Gruber in 2012 can be applied to this setting and offer gains in efficiency relative to MMRM. We use data from the Alzheimer's Disease Neuroimaging Initiative study, to simulate hypothetical clinical trials for a drug that reduces the decline in cognitive impairment among persons with mild cognitive impairment. We compare the performance of MMRM and TMLE for estimating the average treatment effect when varying the prognostic ability of the baseline variables and the models generating patient drop-out.