Analyses

Baseline demographics and ratings were summarized separately for participants who did or did not receive a diagnosis of MCI or dementia during the follow-up period (i.e. after five annual visits where the participant was cognitively normal). We used growth mixture modeling (GMM) ⁴⁶ to model subtypes of NPS trajectories over time. Typically, NPS might be modeled over time using a random effects model ⁴⁷, which would generate estimates of intercept, slope, and possibly quadratic terms to characterize the course of groups of NPS symptoms. GMM allows for the estimation of latent classes (groups) of individuals based on their NPS course, with separate intercept, slope, and quadratic terms generated for each latent class. Since NPI-Q scores were heavily right-skewed, meaning the majority of participants had scores of 0 or close to 0, we used a zero-inflated Poisson (ZIP) model ^48,49. These models are fit in two parts: the first is a logistic model where the outcome is 0 vs >0; the second is a Poisson model with NPI-Q scores greater than 0 treated as count data. To aid model convergence, it is typical to only allow the Poisson part of the model to vary across classes. We used Lo-Mendell-Rubin likelihood ratio tests ^50,51 to choose the optimal number of classes. Model fit was assessed via comparison of observed and predicted trajectories.

We fit Cox proportional hazards models with latent trajectory classes as predictors and adjusting for age at baseline, sex, years of education, and baseline MMSE. We used an adaptation of Vermunt’s three-step method to adjust for uncertainty in class membership ^52,53. In the present case, we used the first four visits of each participant to model the trajectory classes (at least 3 visits are required to plot a curved line), and subsequent visits to model survival. As such, individuals who received diagnoses of MCI or dementia prior to their fifth visit were excluded from the analysis. Had we instead modeled survival using the first four visits as well, we would have introduced circularity; an individual might have been diagnosed with dementia at their second visit, and developed worsening NPS only after that diagnosis. In that case, it would not make sense to say that membership in a worsening NPS class was predictive of increasing risk of dementia, since the temporal ordering had occurred in the reverse direction.

The survival models we fit allow for the inclusion of right-censored individuals (those who do not experience the outcome of interest during the period of observation), but it is assumed that this censorship is not associated with failure ⁵⁴. This assumption of “noninformative censoring” may have been violated in both directions; more impaired volunteers may have had more difficulty attending follow-up visits, but might also be more motivated to be evaluated. This issue is a known limitation of the NACC ⁵⁵ and many other longitudinal studies ⁵⁶.