Nearly every cohort in this book is longitudinal or multi-site or both. PREVENT-AD follows people over years, ADNI and ONDRI scan the same participants repeatedly, and ABCD does it across 21 sites. In all of these, observations are not independent — two scans from one participant resemble each other more than two scans from different participants, and the standard tests assume otherwise.
Mixed-effects models handle this by splitting variance into fixed effects (the population-level relationships you want to estimate) and random effects (systematic variation attached to participants, sites, or scanners). The practical consequence of ignoring the structure is inflated false positives: treating 300 scans from 100 people as 300 independent observations pretends you have three times the information you actually have.
They also handle unbalanced designs gracefully, which matters because real longitudinal cohorts are always unbalanced — people miss visits, drop out, or enrol late. Methods that require complete data force you to discard participants; mixed models use the observations you have.
Related¶
Harmonization — the other way to deal with site
Aging — where longitudinal designs concentrate here