Improving the Reliable Change Index for Identifying Dementia: moving from linear to beta-binomial models

INTRODUCTION: Since linear models are not the most appropriate model for discrete, bounded scores, we compared the beta-binomial (BB) and linear Reliable Chage Index (RCI) to identify progressors to dementia over 13 years among those scoring at the lowest extreme at baseline. METHODS: We calculated the one-year RCI using data from the ASPREE trial (n=15,821), splitting the sample into two halves for cross-validation. Different incidences of dementia (4%, 10%) and standardized mean difference (SMD) between cases and controls (-0.5SD, -0.7SD, -1SD, -1.5SD) were analyzed. Models were compared with the Akaike’s Information Criterion (AIC). RESULTS: Overall, the BB model showed a better AIC than the linear. Across 8,000 random replications, the BB showed a higher positive predictive value in 70%-90% of the cross-validation samples (up to 35% higher), varying for different incidences and SMD. DISCUSSION: The BB model identifies more accurately than the linear model individuals with the greatest risk of progression to dementia.