Reducing the sample size in high‐frequency biomarkers RCTs

• Published in: Alzheimer's & dementia Vol. 16
• Main Authors: Taylor‐Rodriguez, Daniel M, Lovitz, David M, Mattek, Nora, Wu, Chao‐Yi, Kaye, Jeffrey, Dodge, Hiroko H, Jedynak, Bruno Michel
• Format: Journal Article
• Language: English
• Published: 01.12.2020
• ISSN: 1552-5260; 1552-5279
• DOI: 10.1002/alz.042005

Background High‐frequency biomarkers (HFB) are measurements collected weekly, on a daily basis or even more frequently. They are designed to detect early progression from one cognitive state to a later stage. Examples include home‐based measurements, e.g., computer usage and sleep patterns. In randomized controlled trials (RCT), we aim to detect the difference in trajectories of HFB between placebo and experimental group. Current statistical analysis e.g. linear mixed effect with random intercept models detect changes in the mean trajectory but fail to detect changes in the covariance structure, thus requiring longer trials. We examined whether more careful modeling, using historical data, would reduce the sample size needed for a RCT? Objective: (1) to present a novel statistical methodology for analyzing RCTs whose outcomes are HFB based on historical data of cognitively normal subjects, and (2) to evaluate this methodology using simulated data first, then using data from the Oregon Center for Aging and Technology (ORCATECH). Methods The statistical methodology proposed proceeds in three phases (Figure 1). Highlights include the use of historical data for cognitively normal subjects and the use of a Gaussian process with parameterized mean function and covariance structure. Examples of results from simulated data are shown in Figure 2, demonstrating the superior power of the methodology compared to the linear mixed‐effects model. ORCATECH’s weekly computer usage data was used (Figure 3). We computed an RCT effect size with 86 cognitively normal subjects and 11 subjects who were cognitively normal at baseline, but would eventually develop MCI within 139 weeks on average (sd=72). Results Using this new approach, we found an effect size of W=0.616 (Area Under the curve), outperforming a baseline method which yielded an effect size of W=0.515. Respective total sample sizes for an RCT with alpha=0.05 and power=0.8 are 74 and 1028, for the new versus the conventional approach, respectively. Conclusions We have shown that a careful statistical procedure allows for maximally benefiting from HFB by reducing the sample size needed for a 100 week RCT. More experiments will be necessary to validate these findings.