|
Abstract:
|
Within the field of brain imaging, there is an increasing need for longitudinal studies to better answer questions regarding the development and decline of (human) brain function, both in and of itself but also in relation to e.g. cognitive decline or clinical dementia. Technological developments have reduced the price and effort needed to perform brain scans, and more and more longitudinal studies are being planned or performed. In longitudinal studies on human subjects, attrition (dropout) is the rule rather than the exception, and dropout yields missing data. It is reasonable to expect that dropout in brain imaging studies is linked to the brain characteristics of the subject dropping out. Missingness which is associated to the unobserved outcomes is called non-ignorable, or MNAR (Missing-Not-At-Random). In this talk, we will introduce how non-ignorable missingness can be accounted for in the analysis of longitudinal brain imaging studies. We use Bayesian Hierarchical Models, combined with Pattern-Mixture Models, to accomplish this. Our methods are illustrated using real fMRI data.
|
