Background: Analytic models of Alzheimer's disease (AD) tend to focus on one type of symptom and assume implicitly that no measurement error is present. These tendencies render changes in symptom domains difficult to model mathematically, although latent variable methods can accommodate both multiple symptom domains and error. This study formulated and compared underlying (latent) factor structures representing previously reported dependence and independence of symptoms of cognitive decline, functional impairment, and behavioral disturbance in AD.
Methods: In confirmatory factor analyses of data from 2 cohorts of AD patients, 2 levels of latent variables were conceptualized. One general neurologic factor represented disease, and symptom factors represented cognition, function, and behavior. Two "null" models had either a single factor or 3 symptom factors. Two 2-level models treated the general factor as underlying both the observed variables and the symptom factors or treated the symptom factors as explaining variability in the observed variables after taking the general factor into account ("residualized").
Results/conclusions: The residualized model fit the data in both cohorts significantly better than the other models, and relations in this model between some observed and latent variables were different across cohorts. Neither cohort supported a single factor model; both cohorts independently supported a residualized model that may permit differentiation of symptom- from disease-modifying effects of treatment.