Data source
The ADCS-PI design and participants have been described elsewhere [6]. Briefly, the protocol included participants 75 years of age or older who were fluent in English or Spanish. Participants were required to be in good general health and to be on stable doses of all medications for at least 4 weeks prior to screening. Participants were also required to be either cognitively normal or have possible MCI, although in our study we included only those who were cognitively normal (defined by a Clinical Dementia Rating Scale [CDR]-Global score of 0 at baseline) [7]. Participants had to be willing and able to participate in a 4- to 5-year study and be able to complete study procedures. They were also required to enroll with a study partner, with whom they had contact at least twice a week and who would be able to accompany them to all study visits. Participants were excluded if they had medical illnesses that would prevent their participation in the trial, alcohol or substance abuse within the past year, major psychiatric disorders, history of mental retardation, or were participating in a clinical drug trial. The study partner could not have a significant cognitive impairment, based on clinical impression or history known to the investigator, or be enrolled in the study [6].
Participants had annual visits for 4 years after enrollment into the study with a cognitive battery at each visit. Annual cognitive batteries included the modified Mini-Mental State Examination (mMMSE), Free and Cued Selective Reminding Test (FCSRT), New York University Paragraph Recall, Trail Making A and B, Wechsler Adult Intelligence Scale-Revised (WAIS-R) Digit Symbol Substitution Test, Boston Naming Test (10-item version), verbal fluency test (animals), and Geriatric Depression Scale (GDS). At the baseline visit, participants were randomized to take new experimental assessments either at home or in the clinic. All study participants provided written informed consent.
We used the participant and study partner versions of the Mail-In Cognitive Function Screening Instrument, which we refer to simply as the Cognitive Function Instrument (CFI), to compare the information provided by each member of the dyad. The CFI includes 2 sets of questions, 1 each for the participant and the study partner. Each set consists of 14 questions that ask about functional decline in the past year. The study partner and participant were asked to complete these questions independently, but the study partner was allowed to consult other people as long as it was not the participant. Questions were scored as “yes” (1 point), “no” (0 points), and “maybe” (0.5 points). The total score was calculated as the sum of the answers to each of the 14 questions, giving a possible range of 0 to 14. Several questions regarding driving, finances, and work performance had “not applicable” as another possible answer [8, 9]. For questions answered as “not applicable,” we imputed the mean of the available responses.
We used the modified Preclinical Alzheimer’s Cognitive Composite (mADCS-PACC) [9, 10] as an objective measure of cognition. This composite score was a combination of the original mADCS-PACC and the PACC-5 and included total recall from the Free and Cued Selective Reminding Test (FCSRT), New York University Paragraph Recall, Digit Symbol Substitution Test from the Wechsler Adult Intelligence Scale-Revised, mMMSE, and the animal fluency test [10, 11]. For each component, we calculated a z-score using the baseline sample mean and standard deviation for all participants with a CDR-Global score of 0. The z-scores across all components were averaged, and this value was used as the outcome, with lower scores indicating poorer cognition. If any components were missing, the outcome was calculated using the mean of the available components.
Analyses
We used random forests to assess how well each member of the dyad predicted participants’ cognitive state at current and later visits. We used mADCS-PACC score as the response and included the baseline scores for CDR Sum of Boxes (CDR-SB) and participant and study partner CFI along with ethnicity, gender, age, education, and history of cardiovascular disease. Ten observations (0.7%) were missing 1 component of the mADCS-PACC, and 2 observations (0.2%) were missing 3 components. We used the GRF package in R to generate random forests. To account for variability stemming from the generation of the random forests, we created 100 random forests using the same data with different random seeds and obtained a variable importance measure each time. The variable importance measure used was a weighted sum of the number of times variables were selected for splitting, with splits higher in the tree having larger weight. We calculated the mean of the variable importance estimates, which we refer to as the estimated mean variable importance (eMVI), along with corresponding 95% uncertainty bounds (UB), which were obtained using the 2.5th and 97.5th percentiles of the 100 observed variable importance values. To investigate whether participant cognitive performance affected the observed relationships, we stratified participants using quartiles of the mADCS-PACC at baseline. The lowest quartile included participants with the poorest cognitive performance while the highest quartile included participants with the highest cognitive performance. To quantify differences in predictive performance between study partner types, we further stratified participants into spousal and non-spousal dyads. We also fit linear regression models for each stratum adjusting for the same variables that were used to generate the random forest results. To investigate how well current participant and study partner CFI scores predict current cognitive state, we repeated this procedure for each of the 4 years using cross-sectional scores for the mADCS-PACC, CFI, and CDR-SB.
Fifteen observations were excluded from the baseline analysis because either baseline participant or study partner CFI was missing. Seventy-two observations were excluded from the cross-sectional analysis because cross-sectional participant or study partner CFI was missing. Participants who dropped out were asked to return for a final visit. These visits were included in our analyses as the closest annual visit. Two treatment discontinuation visits were removed from our analysis because the participants already had a visit for the year closest to the treatment discontinuation. Three participants who were missing mADCS-PACC at baseline were also excluded from the cross-sectional analysis.
