Joint trajectories of episodic memory and odor identification in older adults: patterns and predictors

Characteristics of the study population

The characteristics of the 1023 subjects are shown in Table 1. The majority, 794 (77.6%), were female, and the mean age at baseline was 78.2 (SD 7.5) years. The mean B-SIT and episodic memory (z-score) were 9.4 (SD 2.0) and 0.3 (SD 0.6) at baseline, respectively. During the follow-up period, among all participants, 290 (28.3%) participants died, and the participation rate of survivors exceeded 90%. The median number of assessments was 8 (IQR=8-5).

Heterogeneity of trajectories

Quadratic curves were fit for episodic memory and odor identification trajectories separately and in a joint model across 8 years. The observed individual trajectories of episodic memory and B-SIT means of the entire sample are depicted in Supplementary Figure 1A, 1E. The trajectories of B-SIT and episodic memory were clearly related, as shown by the positive correlation between their random slopes (1-class model; r= 0.57, p < .001). The joint trajectories mean of the entire sample is depicted in Figure 1A.

Figure 1. Fitted and observed trajectories of jointly modelled episodic memory and B-SIT scores. Scales are z-scores (mean 0, SD 1). (A) The trajectories of the entire sample (n= 1023; (B) Latent Class 1-Stable average (n= 690, 67.4%); (C) Latent Class 2- Stable EM/OI decline (n= 231, 22.6%); (D) Latent Class 3- Joint decline (n= 102, 10.0%). The mean trajectories of each plot are shown in bold.

Episodic memory

When fitting models with increasing numbers of classes, the 3-class model provided the best balance between model fit and model complexity, as confirmed by the LMR test (3- vs 4-class model: −2LL(4) = 1895.3, p = 0.321). An overview of the model fit criteria is shown in Supplementary Table 1. The difference in BIC between the 3-class model and the 4-class model was small, indicating that the model fit improvement caused by the 4^th class was minimal. The best-fitting 3-class model included a class-specific intercept variance and class-invariant slope variance. The parameter estimates of the 3-class model are shown in Supplementary Table 2, and the trajectories are depicted in Supplementary Figure 1B–1D. Class 1 was the largest class (n=799, 78.1%) and was characterized by above average episodic memory at baseline, and a stable trajectory over time. Class 2 was the second largest class (n=158, 15.4%), showing average episodic memory at baseline, as well as a quadratic decline over time. Class 3 was the smallest class (n=66, 6.5%), with below average episodic memory at baseline and faster decline over time.

Odor identification

When we fitted models with increasing numbers of classes, the 3-class model provided the best fit according to the LMR test (3- vs 4-class model: −2LL(4) = 358.48, p = 0.095) and the class sizes (Supplementary Table 1). The difference in BIC between the 3-class model and the 4-class model was small, indicating that the model fit improvement caused by the 4^th class was minimal. The best-fitting 3-class model included a class-specific intercept variance and class-invariant slope variance. The parameter estimates of this 3-class model are shown in Supplementary Table 1 and the trajectories are depicted in Supplementary Figure 1F–1H. Class 1 was the largest (n= 731, 71.5%) and showed above average odor identification at baseline, which remained stable over time. Class 2 was the smallest class (n=79, 7.8%), showing around average odor identification performance at baseline, as well as a linear decline over time. Class 3 was the second largest (n=212, 20.7%), with the lowest odor identification at baseline, remaining rather stable over time.

Episodic memory and odor identification

When we fitted models with increasing numbers of classes, the 3-class model provided the best balance between model fit and model complexity, as confirmed by the LMR test (3- vs 4-class model: −2LL(7)=908.95, p=0.345 (Supplementary Table 1). The difference in BIC between the 3-class model and the 4-class model was small, indicating that the model fit improvement caused by the 4^th class was minimal. The 3-class model with the best fit included a class-specific intercept variance and class-invariant slope variance. The parameter estimates of this 3-class model are shown in Table 2, and the trajectories are depicted in Figure 1B–1D. Participants in Class 1 (690, 67.4%) exhibited joint stable average performance, those in Class 2 showed stable average episodic memory and decline in odor identification (n=231, 22.6%), and Class 3 was characterized by joint decline in both episodic memory and odor identification (n=102, 10.0%).

The sample characteristics of the three classes are shown in Table 1. Noteworthy is that 144 out of 690 (20.9%) participants died in Class 1 (stable average) during follow-up, 84 out of 231 (36.4%) died in Class 2 (average episodic memory/declining odor identification), and 62 out of 102 (60.8%) participants died in Class 3 (joint decline); 59 participants (8.6%) developed dementia over follow-up in Class 1, 59 (25.5%) in Class 2 and 69 (67.7%) in Class 3; participants in Class 1 were younger than those in Class 2 or 3, however the latter two classes did not differ significantly in age.

Predictors of class membership

Having established 3 distinct patterns of decline and their relative occurrence, we next sought to determine what predictors could be used to characterize these groups. We first predicted decline outcomes for episodic memory, then for odor identification, and finally for both functions.

Episodic memory

Using multivariable logistic regression, potential predictors of class membership listed in Table 1 were examined, with the class membership from the final 3-class model of episodic memory the dependent variable. This analysis was based on 914 participants; 109 (10.7%) were excluded due to missing values in covariates; the excluded participants did not differ significantly in B-SIT or episodic memory at baseline. Higher age at baseline and being an APOE ε4 carrier increased the likelihood of belonging to the lower performing Classes 2 and 3, relative to Class 1 (stable class). Higher BMI and cognitive activity decreased the likelihood of being in Class 3 relative to Class 1 (Supplementary Table 3).

Odor identification

In the multivariable logistic regression, with the class membership from our final 3-class model of odor identification as the dependent variable, higher age at baseline was associated with higher odds of belonging to Class 2 relative to Class 1 (stable class). Moreover, higher baseline age and being an APOE ε4 carrier increased the likelihood of belonging to the lowest performing Class 3 relative to Class 1 (Supplementary Table 3).

Episodic memory and odor identification

Table 3 shows the significant predictors of joint class membership presenting three patterns (joint stable, declining odor identification, and joint decline). Compared to the stable Class 1, people in the declining odor identification Class 2 were more likely to be older and male. People in the joint declining Class 3 were more likely to be older, APOE ε4 carriers, have a lower BMI, and be less engaged in cognitive activities (Table 3).

Sensitivity analysis

As both decline in episodic memory and olfaction are strongly related to aging, we conducted a sensitivity analysis by rerunning the models of joint prediction of class membership for episodic memory and odor identification, adjusting for age at baseline. The best-fitting 3-class model included a class-specific intercept variance and class-invariant slope variance.

The parameter estimates of the 3-class model are depicted in Supplementary Table 4. Participants in Class 1 (747, 73.0%) exhibited joint stable average performance, those in Class 2 showed stable average episodic memory and decline in odor identification (n=216, 21.1%), and Class 3 was characterized by joint decline in both episodic memory and odor identification (n=60, 6.0%).

Study population

The Rush Memory and Aging Project (MAP) is an ongoing prospective clinical-pathologic study on risk factors for common chronic conditions of old age. Eligibility for participation involves agreeing to annual clinical evaluations and to the donation of brain, spinal cord, and selected nerves and muscles to Rush investigators at death. Detailed information on the MAP study design and the evaluation protocol is provided elsewhere [39]. In brief, participants were recruited from church groups, senior centers, retirement communities, and senior citizen housing facilities, within the greater Chicago area. At baseline and thereafter, all participants underwent extensive clinical assessments, including detailed medical history, neurological examination, extensive cognitive function testing, and odor identification testing.

Enrollment began in 1997. However, annual olfaction testing did not start until 2011, therefore 2011 serves as the analytic baseline for these analyses. Through 2018, 1041 participants were enrolled and annually followed-up. Out of the 1041 participants, we excluded participants with prevalent dementia (n = 18), leaving 1023 for the current analyses.

The study was approved by the Institutional Review Board of Rush University Medical Center and was performed in accordance with the ethical standards laid down in the 1964 Declaration of Helsinki and its later amendments. Informed consent was obtained from all participants. All participants signed an Anatomic Gift Act for organ donation. Participants also signed a repository consent that allowed their data to be shared. More information on obtaining data can be found on the RADC Resource Sharing Hub at https://www.radc.rush.edu.

Data collection

Participants underwent uniform evaluations with trained staff including structured interviews, clinical and neurological examinations, and cognitive testing as described previously [39]. Data on socio-demographic characteristics (i.e., age, sex, and education), lifestyle factors (i.e., smoking), medical conditions, and cognitive function were collected at each wave following standardized procedures [39].

Education was recorded as participants’ maximum years of formal schooling. Smoking was categorized as “never smoked”, “former smoker” and “current smoker”. Self-reported information on medical conditions including heart disease, hypertension, diabetes was collected during the interview at baseline [39]. at study entry blood samples were taken and the APOE gene was genotyped utilizing high-throughput sequencing. Participants were stratified as epsilon 4 (ε4) carriers or ε4 non–carriers. Depression was determined according to the criteria of the Diagnostic and Statistical Manual of Mental Disorders, 3rd ed, Revised, implemented with a subset of questions from the Diagnostic Interview Schedule at baseline [40]. Clinical diagnosis of dementia was conducted on the basis of criteria of the joint working group of the National Institute of Neurological and Communicative Disorders and Stroke and the Alzheimer’s disease and Related Disorders Association (NINCDS/ADRDA) [41].

Outcomes

Odor identification

At the analytic baseline and at each follow-up, the Brief Smell Identification Test (B-SIT) (Sensonics. Inc., Haddon Heights, USA) was administered. The B-SIT is a 12-item standardized test, with 4 alternatives per item. In this test, a booklet is presented containing a scratchable patch of microencapsulated odorant on each page. The examiner scratches the odor patch with a pencil to release the odorant and places it under the participant’s nose, the participant is then asked which of four specific odors the sample most closely resembles. The score reflects the number of odors correctly recognized, with possible scores ranging from 0-12. If an item response is missing for a maximum of two, a score of 0.25 is assigned, corresponding to a chance level performance. If responses to three or more items are missing, data on this test are considered missing. The B-SIT has previously shown to be internally consistent, with good correspondence to the 40-item University of Pennsylvania Smell Identification Test from which it was derived [27].

Episodic memory

Episodic memory was measured using CERAD (Consortium to Establish a Registry for Alzheimer’s Disease) Word List Memory, Recall, and Recognition, immediate and delayed recall of the East Boston Story, and Logical Memory immediate and delayed recall. Raw scores on each test were converted to z-scores (i.e. standardized based on all MAP participants at baseline) and then the average z-score among the tests for episodic memory was computed, as reported in detail in a previous study [28].

Assessment of lifestyle factors

Three measures of stimulating mental, social and physical activities were recorded during the baseline interview.

Cognitive activities

Participants completed a 7-item cognitive activity questionnaire [22, 36]. Activities include reading, writing letters, visiting a library, and playing games such as chess or checkers. Frequency of participation in each activity was rated from 1 (once a year or less) to 5 (every day or about every day) [36].

Social activity

The frequency of activity was assessed on a 6-item scale including: a) going to restaurants or sporting events; b) going on day or overnight trips; c) doing unpaid community or volunteer work; d) visiting relatives’ or friends’ houses; e) participating in groups, such as a senior center; and f) attending church or religious services. Participants were asked to rate how often they participated in each activity based on a 5-point scale from 1 (once a year or less) to 5 (every day or about every day) [36].

Physical activity

Frequency of activity was recorded as the hours per week participants reported engagement in 5 categories of activities: walking for exercise, gardening or yard work, calisthenics or general exercise, bicycle riding (including stationary bikes), and swimming or water exercises [36].

Predictors on class membership

The following variables were included as potential predictors of class membership: age (years); sex (male/female); education (years); history of hypertension (yes/no), depression (yes/no), smoking (never/previous/current,) APOE ε4 status (any ε4/no ε4), diabetes (yes/no), history of heart failure (yes/no), BMI (continuous), social activity (hours per week), cognitive activity (hours per week), physical activity (hours per week).

Statistical analyses

We used growth mixture models (GMMs) to model trajectories of episodic memory and odor identification separately and jointly over time. We used yearly assessments of episodic memory and odor identification to model the trajectories. The time scale was time from the first assessment (2011) until last follow-up (2018), maximum 8 years. Episodic memory and odor identification were assessed yearly; thus the timescale is in years. The GMM is a longitudinal form of latent class analysis, using mixed models. The GMMs therefore groups participants into latent classes, on the basis of similarities in their trajectory patterns over time. This is done by fitting an increasing number of curves until an optimal balance between model fit and model complexity is reached. Quadratic models with 1 to 5 classes were fit, and the final model was chosen based on the Bayesian information criterion (BIC), Lo-Mendell-Rubin (LMR) likelihood ratio test, and class size. The BIC indicates the fit of a model, the lower the value, the better the fit of the model is. The LMR test is used to compare model fit between 2 nested models. A significant LMR test indicates that the model with k classes has a better fit than the same model with k − 1 classes. The parameter estimates were obtained using maximum likelihood estimation, with standard errors (SEs) that are robust to non-normality. The quadratic slope variance was fixed to zero. For ease of interpretation and presentation, the B-SIT scores were standardized (based on all participants at baseline) for comparability with the global cognition z-score. For parsimony, the residual variances assumed to be equal across classes and were allowed to vary over time. After the model with the best fit was decided, we examined which factors predicted class membership using multinomial logistic regression with the 3-step method in a multivariable model [42]. The GMMs were fit using Mplus version 8.2. Further analyses and processing of results and multinomial logistic regression models, were performed using Stata v. 15 and RStudio v. 1.2.5001.