Abstract
OBJECTIVES: Epidemiologic models used to model cognitive aging and decline often rely on a linear random slopes model, although clinicians rely on accelerated cognitive decline for diagnosis. This study examined the extent to which accelerated versus linearized models of cognitive decline are biased by study design and analytic choices. METHODS: Data on participants of the Health and Retirement Study (1998-2019) who completed cognitive monitoring and lacked cognitive functional limitations or a history of stroke at baseline were included. Episodic memory was measured as the outcome at each time-point. Experiments manipulated sample size, gaps between observations, and waves/participants, and the degree of censoring due to factors in the model, and the degree of cognitive functional limitations. The outcome was the degree of bias in rates of cognitive decline estimated by different models when compared with the full sample. RESULTS: This study used data from 30,740 Health and Retirement Study participants followed 194,818 times and for 335,025.73 person-years of cognitive assessment. While linear models were often biased by relatively small changes in sample design, attrition, and analytic choices, quadratic models and nested nonlinear models provided results that were less biased. Nested nonlinear models were less biased than linear models across 4/5 experimental conditions and were less biased than quadratic models in 3/5 experimental conditions. DISCUSSION: Linearized models of cognitive decline yielded high levels of bias that were sensitive to variations in study design and analytic choices. Results support shifting toward methods modeling accelerated decline in studies of cognitive aging and decline.