Statistical models for Alzheimer's disease clinical trials: Lessons learned from the DIAN-TU Platform Trial

Abstract

Background

Alzheimer's disease (AD) clinical trials often involve uneven follow-up durations and long-term open-label extensions (OLE), yet conventional statistical models are typically designed for fixed schedules, limiting their efficiency in such settings.

Objective

To describe and illustrate alternative statistical modeling approaches developed and implemented in the Dominantly Inherited Alzheimer Network Trials Unit platform trial to optimally leverage data with irregular and extended follow-up.

Methods

We present three complementary models: (1) a Cox proportional hazards model for recurrent disease progression events that uses all observed worsening events rather than only the first event; (2) a parametric disease progression model based on estimated years from expected symptom onset that estimates proportional slowing or time delay in disease progression; and (3) piecewise linear mixed-effects models tailored to the “gap” period between the double-blind phase and OLE, accommodating variable off-treatment intervals and missing interim data. All methods are illustrated with hypothetical examples, and ready-to-use SAS code is provided in the Supplemental Material.

Results

The proposed models successfully handle complex longitudinal data structures typical trials with OLE phases, offering greater statistical efficiency and more comprehensive capture of treatment effects over extended periods compared with traditional approaches.

Conclusions

These flexible, efficient statistical models are well-suited for rare disease and long-duration AD trials. Wider adoption and further validation of these approaches may enhance the power and interpretability of future neurodegenerative disease trials.