Joint models for estimating determinants of cognitive decline in the presence of selection bias after enrollment
Embargo Date
2023-02-17
OA Version
Citation
Abstract
Alzheimer’s disease and related dementias are a public health burden because declining cognitive function is associated with adverse health outcomes. Despite decades of research, highly effective preventive strategies remain elusive. Several methodological problems arise in studies of cognitive decline such as selection bias which can lead to spurious associations or reverse the direction of association. We focus on evaluating and developing joint models for estimating determinants of cognitive decline in the presence of selection bias after enrollment, such as informative drop-out and visitation times.
Joint models for longitudinal and time-to-event data combine a linear mixed effects model (LME) with a Cox proportional hazards model to simultaneously estimate cognitive decline and drop-out. Studies have shown that correctly specified joint models produce unbiased estimates. However, no studies have evaluated the robustness of joint models to misspecification of the drop-out process. We utilize an existing simulation platform to evaluate the robustness of commonly used modeling frameworks: generalized estimating equations (GEE), weighted GEE, LME, and joint models in estimating the effect of an exposure on cognitive decline.
Further, depending on the study design, participants experiencing cognitive decline might be more or less likely to be observed making the visitation process informative. Joint models for longitudinal, repeated, and terminal events exist, but they do not allow for an association between slope in cognition and recurrent or terminal event and assume the frailty term follows a lognormal distribution instead of gamma. We propose a novel joint model for estimating determinants of cognitive decline in the presence of informative visitation and drop-out and assess our model under various levels of each process.
In many longitudinal studies of cognition, it is of interest to identify determinants of pre-dementia cognitive decline. Therefore, follow-up for cognitive measures could stop prior to the end of study due to dementia diagnosis or death. We extend our novel joint model to include a competing terminal event and assess the sensitivity of the longitudinal estimates. Finally, we include an application of our novel joint model using neuropsychological data from the Framingham Heart Study to evaluate the effect of higher education on pre-dementia cognitive decline.