Reproducibility of applying multiple statistical analyses to data from the TRAILBLAZER‐ALZ clinical trial of donanemab in early symptomatic Alzheimer's disease

Abstract

AbstractBackgroundThe Mixed Model for Repeated Measures (MMRM) is the most frequently used statistical analysis for clinical trial endpoints in Alzheimer’s disease (AD) but has the limitation of not efficiently leveraging data from intermediate time points due to the flexibility in the mean structure. In the TRAILBLAZER‐ALZ trial, donanemab, a humanized IgG1 antibody specifically targeting brain amyloid plaque, slowed disease progression in individuals with early symptomatic AD. The aim of this presentation is to apply different statistical models to clinical scale data gathered from this trial, as consistent conclusions following multiple methods would strengthen confidence in the results.MethodTRAILBLAZER‐ALZ (NCT03367403) was a randomized, placebo‐controlled, double‐blind, multi‐center Phase 2 study assessing the safety, tolerability, and efficacy of donanemab in patients with early symptomatic AD. The change from baseline to 76 weeks was assessed using the Integrated AD Rating Scale (iADRS), a composite tool measuring cognition and daily function (primary endpoint), and the Clinical Dementia Rating Scale‐Sum of Boxes (CDR‐SB; secondary endpoint). In addition to MMRM, three statistical models were fit: Bayesian Disease Progression Model (DPM), Natural Cubic Spline (NCS) model, and Quadratic Mixed Model (QMM). Each model has unique assumptions and data handling, for example, in treating time categorically or continuously, which can lead to differences in power while still controlling the false‐positive rate to an allowable margin.ResultThe percent slowing estimates and treatment differences at 18 months from the models are consistent in supporting a positive treatment effect across the clinical scales. The DPM, NCS, and QMM provide a reduction in the standard error, which leads to tighter intervals due to the model assumptions generally being met.ConclusionThe consistency of the results between the statistical models and across the clinical scales suggests the general robustness and strength of the data, regardless of the statistical model used. DPM, NCS, and QMM may provide a more detailed estimate of the longitudinal trajectory of each treatment group when there are small sample sizes and noisy data, and an MMRM model may not be an ideal choice in trial designs that may have sparse data at later time points.