Our analyses included the first 1,135 consecutive participants. We excluded 86 persons clinically rated as having dementia at baseline. Demographic characteristics are reported in Table . The average baseline age was 75 years, the sample was predominantly white and over 90% completed greater than 12 years of education. The number of evaluations ranged from 1 to 16 and the median length of follow-up was 8 years. The overall follow-up rate exceeded 95%. At the time of these analyses, 422 persons had died, of whom 395 had a brain autopsy. Summary data on amyloid and tangle pathology was available for 281 participants.
Participant characteristics at first observation (n = 1,049)
The pace of change per year was −0.15 SD units [confidence interval (CI): −0.17 to −0.14] in the single class model. Given the baseline total sample SD, participants would be expected to experience a half SD decline in performance in about 3.3 years, likely representing a noticeable difference in cognitive performance [22
Parameter estimates for the three class model are shown in Table . The first class (slow decline) included 678 participants and was characterised by a mean composite score at baseline of 0.17, implying that this class had a baseline age-adjusted mean composite score slightly above the total sample baseline average. Linear change over age was slow (slope = −0.04 SD units per year of age, CI: −0.05 to −0.03). A second class (moderate decline; n
= 284) had a baseline composite score of −0.17 (about 0.2 SD below the overall sample baseline average). This class had a faster pace of change with age (slope = −0.19, CI: −0.22 to −0.16), declining about five times as fast as the slow decliners and experiencing a half SD decrement in cognitive performance in about 2.6 years. A third class (fast decline, n
= 87) had low baseline scores (initial level = −0.32) and a fast rate of annual decline (slope = −0.57, CI: −0.65 to −0.48). On average, members of this class would experience a half SD decline in about 11 months. Retest effects differed by trajectory class. The fast decline class had a higher retest effect between their first and second observation, an average of 0.74 SD units, relative to 0.40 units in the moderate class and 0.27 units in the slow decline class. The Supplementary data available in Age and Ageing online, Appendix 2 figure
shows plots of model-implied cognitive trajectories over age by class and age groups.
Parameter estimates from random effects mixture model of change of a composite global cognitive functioning score over time (up to 15 years; n = 1,049)
The overall empirical r2 for the three class model was 0.94. Fit was better at baseline for the slow and moderate classes. Posterior probabilities summarise the probability of membership in each class and were highest for the fast class (0.96). Entropy was 0.73; generally, values greater than 0.80 indicate good classification. Therefore, this model may constitute an over-extraction of classes.
For comparison, results from the two class model identified slow (n = 799) and fast decline classes (n = 250). The initial level of the slow class was +0.11 with an annual pace of change of −0.06. The initial level for the faster class was −0.25 with an annual pace of change of −0.38. The two class model had an entropy value of 0.81, but, had a higher aBIC value and a lower log-likelihood than the three class model. Both the two and three class models had an empirical r2 = 0.94. Therefore, the evidence suggested the three class model fit better. We retained the three class model because differences were quantitative rather than qualitative and our analytic goal was description and modelling, rather than development of a classification tool. We investigated model modifications for possible misspecification, including the assumption of homoscedasticity and the common effect of age group indicators on baseline scores and change over time across class. Modifications did not significantly improve model fit.
Study participants differed significantly by class at baseline on age, education and Mini Mental State Examination (MMSE) [23
] score (Table ). Relative to the slow decline class, the fast decline class had higher cumulative risk of death (relative risk (RR) 2.2, CI: 1.9 to 2.6) as did the moderate decline class (RR 1.5, 95% CI: 1.3 to 1.8). The fast decline class had the highest relative frequency of people with one or more APOE
ε4 allele(s) (31/87 = 39%), about twice that in the slow decline class (114/435 = 21%).
Participant characteristics at first observation by most likely class membership (n = 1,049)
In the autopsy subsample (n
= 281), the slow, moderate and fast decline classes were not significantly different in age, sex or education levels. Moderate and fast decline groups had more visits (χ2
= 6.0, P
= 0.05) and longer follow-up (χ2
= 7.4, P
= 0.03) than the slow decline group. We found significant differences in amyloid burden and tangle density across classes, amounting to a medium effect size [24
] for amyloid load (Cohen's f
= 0.25) and large effect (Cohen's f
= 0.36) for tangle density (see Supplementary data available in Age and Ageing online