Cognitive assessments in ADNI: Lessons learned from the ADNI psychometrics project

Abstract

AbstractBackgroundThe Alzheimer’s Disease Neuroimaging Initiative includes cognitive assessments at every study visit. We have had the opportunity to use modern psychometric approaches including confirmatory factor analysis (CFA) and item response theory (IRT) with these rich data. Here we discuss lessons learned from analyses of memory, executive functioning, language, and visuospatial functioning in ADNI.MethodWe analyzed data from all ADNI study waves. We considered granular data for each test. Our panel of experts considered each granular data element and assigned them to a single primary domain – memory, executive functioning, language, or visuospatial functioning – or to none of these. We then used CFA and IRT approaches to calibrate each domain separately. ADNI has extensive imaging and fluid biomarker data we used for validity assessments.ResultThe ADNI cognitive battery emphasizes memory assessment with multiple indicators. The assessment of language and executive functioning are less robust but are each measured with several indicators. The assessment of visuospatial functioning is much sparser. Assessment intensity is nicely reflected in measurement precision data, where standard errors of measurement are smallest for memory, intermediate for language and executive functioning, and largest for visuospatial functioning. Validity assessments were solid for all four domains. We have uploaded resulting composite scores (ADNI‐Mem, ADNI‐EF, ADNI‐Lan, and ADNI‐VS) to the Laboratory On Neuroimaging (LONI)‐hosted ADNI website.ConclusionThe composite scores we have developed will be useful for investigators integrating domain‐specific cognitive data in their analyses of ADNI data. The measurement precision findings have implications for future study design; for example, more robust measurement of visuospatial functioning may be a wise investment. Measurement precision findings also have implications for analytic strategies. Standard typical analyses ignore measurement precision and treat observed scores as if they were measured without error. We are developing hierarchical Bayesian modeling strategies that can incorporate both point estimates for domain scores along with standard errors of measurement. These models will account for measurement error, ensuring that inferences are valid across domains with very different measurement properties. We will make this code widely available as well.