Background: Accurate longitudinal modelling of cognitive decline is a major goal of Alzheimer's disease and related dementia (ADRD) research. However, the impact of subject-specific effects is not well characterized and may have implications for data generation and prediction.
Objective: This study seeks to address the impact of subject-specific effects, which are a less well-characterized aspect of ADRD cognitive decline, as measured by the Alzheimer's Disease Assessment Scale's Cognitive Subscale (ADAS-Cog).
Methods: Prediction errors and biases for the ADAS-Cog subscale were evaluated when using only population-level effects, robust imputation of subject-specific effects using model covariances, and directly known individual-level effects fit during modelling as a natural control. Evaluated models included pre-specified parameterizations for clinical trial simulation, analogous mixed-effects regression models parameterized directly, and random forest ensemble models. Assessment used a meta-database of Alzheimer's disease studies with validation in simulated synthetic cohorts.
Results: All models observed increases in variance under imputation leading to increased prediction error. Bias decreased with imputation except under the pre-specified parameterization, which increased in the meta-database, but was attenuated under simulation. Known fitted subject effects gave the best prediction results.
Conclusion: Subject-specific effects were found to have a profound impact on predicting ADAS-Cog. Reductions in bias suggest imputing random effects assists in calculating results on average, as when simulating clinical trials. However, reduction in error emphasizes population-level effects when attempting to predict outcomes for individuals. Forecasting future observations greatly benefits from using known subject-specific effects.
Keywords: Alzheimer’s disease; cognitive dysfunction; forecasting; mental status and dementia tests; statistical models.