Analytical Choices of Timescale in the Presence of Practice and Cohort Effects in Longitudinal Cognitive Aging Research: A Simulation Study Using Linear Mixed‐Effects Models
Jingxuan Wang, Ruijia Chen, Sarah F Ackley, Scott C Zimmerman, Kendra D Sims, Eleanor Hayes‐Larson, Elizabeth Rose Mayeda, M. Maria Glymour- Psychiatry and Mental health
- Cellular and Molecular Neuroscience
- Geriatrics and Gerontology
- Neurology (clinical)
- Developmental Neuroscience
- Health Policy
- Epidemiology
Abstract
Background
Understanding determinants of cognitive decline relies on several modeling assumptions. Alternative longitudinal modeling strategies–such as accounting for aging using time‐on‐study (estimated based on within‐person comparisons), baseline age (estimated based on between‐person comparisons), or current age (estimated based on both within‐ and between‐person comparisons) –may produce qualitatively different results. Using simulations, we evaluated robustness of effect estimates for predictors of cognitive decline with current age (vs. time‐on‐study) as the timescale, in the context of practice and/or cohort effects.
Method
Our simulation strategy approximates the data structures of several cohort studies of cognitive aging (figure 1). For scenarios with only practice effects and with both practice and cohort effects, we performed 1,000 simulations of a cohort of 2,000 individuals with annual cognitive assessments over 3 years. Longitudinal models were fit using three timescale specifications (with and without adjustment for practice effects for each): (1) time‐on‐study adjusted for baseline age; (2) current age without adjustment for baseline age; and (3) current age adjusted for baseline age. We evaluated the bias in estimated effect associated with each specification by comparing to known simulation parameters.
Result
When practice effects were present and not accounted for, time‐on‐study and current age underestimated rate of cognitive decline (figure 2) but bias was smallest for current age as the timescale. When cohort effects were present, current age as the timescale led to biased estimates when baseline age was not adjusted for. With a correctly specified model, the variance in the estimate of the coefficient for current age was 36.3% smaller than the variance in the estimate of the coefficient for time‐on‐study with baseline adjustment. All estimates for the effect of education on cognitive decline were unbiased regardless of timescale, but 95% CIs were narrowest for current age as the timescale.
Conclusion
In this simulation, current age as the timescale that combines within‐ and between‐person estimates led to more precise estimates than time‐on‐study with baseline age adjustment. Failure to account for practice effects can bias estimates of longitudinal within‐person age effects.