Slower Decline in Processing Speed Is Associated with Familial Longevity

Intact cognitive function is vital to all higher order functioning in daily life and thus is a cornerstone of healthy aging. Numerous studies have shown that cognitive impairment is associated with a greater risk of dementia [1], disability [2, 3], and mortality [4]. As more than 1 in 5 older adults over the age of 70 years in the USA is cognitively impaired [5], it is more important than ever to identify groups of people who are able to delay or avoid cognitive impairment so that we can identify protective factors and underlying mechanisms associated with the preservation of good cognitive function throughout life.

Individuals who are diagnosed with mild cognitive impairment (i.e., modest declines in cognition with retained ability to perform everyday activities) or dementia (i.e., substantial declines in cognition resulting in impaired ability to perform everyday activities [6]) show an accelerated decline across one or more domains of cognitive function beginning many years prior to diagnosis [7, 8]. Yet even in normal aging, some cognitive functions, particularly those requiring speed and manipulation and processing of information are known to show decline within the range of normal cognitive function [see [9] for review, [10]]. In the Whitehall II study individuals aged 45–70 years showed declines in short term memory, reasoning, and verbal fluency over 10 years whereas vocabulary increased or remained stable on average over time [11]. This is of concern because cognitive decline is associated with an increased risk of mortality [4, 12] suggesting an opportunity for intervention. In studies investigating the relationship between cognition and survival among individuals without dementia, poorer cognitive function in the domains of executive function (e.g., abstract reasoning and processing speed), visuospatial ability (e.g., visuoperception, visuoconstruction, and visual retention), and language (e.g., naming, verbal fluency, production), were associated with a higher risk of mortality [13, 14]. However, there is substantial between-person variability in trajectories of cognitive change [15]; Wilson and colleagues [16] found that only a portion of nondemented older adults showed declines in cognitive function, whereas some remained stable or even improved over time. There is also evidence that between-person variability may increase with increasing age, particularly in the domains of memory and executive function [17, 18], as risk of cognitive decline increases in association with a higher risk of mortality and underlying neurodegenerative disease [19]. Thus, the identification of individuals who maintain stable cognitive function or exhibit slower cognitive aging (i.e., less decline over time) would allow for the investigation of factors associated with healthy cognitive aging, and conversely, the identification of markers of increased risk of mortality and neurodegenerative disease.

Although age is the greatest risk factor for cognitive decline, 30–55% of individuals at the most advanced ages (i.e., centenarians) are able to avoid dementia [20, 21]. Yet centenarians are at the end of their lives and have a high mortality rate, which hinders the ability to study their longitudinal trajectories of cognition and life course factors that affect cognitive decline. Since longevity and longer health spans tend to run in families [22, 23], and cognitive function can be heritable [24], studies have found that offspring of long-lived parents (i.e., individuals with a parent aged 85+ years) show better cognitive function, a lower risk of dementia, slower cognitive decline, and less change in ventricle size than their peers (i.e., individuals from the same birth cohort without evidence of familial longevity) [25-27].

Family-based studies confer added power for identifying shared genetic and environmental factors that contribute to health and function outcomes. Therefore, studying families enriched for longevity may offer insights into the processes that confer resilience to cognitive aging that differ from those that could be identified among offspring of sporadic parental longevity (e.g., centenarian offspring). The Long Life Family Study (LLFS) is a two-generation study of families throughout North America and Denmark who were selected for clustering of longevity. Many LLFS family members have longer health spans and age better than their peers (i.e., reference populations not selected for longevity) as evidenced by lower disease prevalence, later disease onset, and better physical function [22, 23, 28]. They also have better cognitive function than individuals without familial longevity. For example, the proband generation (average age 91 years at enrollment) was found to have better global cognitive function than participants in the Cardiovascular Health Study (CHS) and the Framingham Heart Study, cohorts that were not selected for longevity [28]. In the offspring generation (average age 61 years at enrollment), family members had better processing speed than CHS participants (no differences in global cognitive function most likely due to ceiling effects of the Mini-Mental State Examination). Additionally, offspring of the oldest probands (i.e., nonagenarians and centenarians) had a 60% reduced odds of cognitive impairment consistent with probable Alzheimer’s disease compared with their spouses at an average age of 70 years [29]. They also outperformed their spouses on tests of attention, working memory, and fluency but showed no differences in episodic memory, after taking into account differences in age, sex, education, and health status [30]. Yet, it is unknown whether the lower prevalence of cognitive impairment and superior cognitive function across multiple tasks among individuals with familial longevity primarily reflects lifelong differences in cognitive functioning that persist into late life, or whether this cognitive advantage also reflects a slower rate of cognitive aging. If cognitive advantages associated with familial longevity are due to underlying lifelong differences in brain structure or function, we would expect to see parallel trajectories of change on cognitive tests between members of long-lived families and spouses with a sustained superior performance of family members on tests of processing speed, attention, working memory, and fluency. However, if familial longevity also confers resilience to cognitive aging, trajectories of cognitive change between family members and spouses will diverge as family members show a slower rate of change over longitudinal assessments in comparison to spouses.

LLFS family members and their spouses underwent comprehensive cognitive testing at enrollment, and 67% of them repeated the assessment approximately 8 years later, allowing for investigation of rates of change in specific aspects of cognition over time. The cognitive assessment included tests of global cognitive function, auditory attention, working memory, episodic memory, semantic fluency, and psychomotor processing speed. The primary aim of the current study was to identify baseline differences in cognitive function between LLFS family members and their spouses and determine the extent to which rate of cognitive decline (rate of change per year) differs between them using longitudinal measurements of 6 neuropsychological test scores. We hypothesized that individuals from long-lived families have longevity-associated factors (e.g., favorable genetics, lifestyle habits, and environmental contributions) that confer higher baseline cognitive function in processing speed, attention, working memory, and fluency and reduce the rate of cognitive decline compared with their spouses in the domains of processing speed and episodic memory which decline with aging and neurodegenerative disease.

The LLFS is a multicenter study of familial longevity and healthy aging. Three field centers in the USA located in Boston, NY, and Pittsburgh and 1 field center in Denmark enrolled 5,086 participants from 592 families during the period 2006–2017. Families were chosen based on the Family Longevity Selection Score [31], which calculated the survival probabilities of siblings in the proband generation of each family. Families were required to have at least a living sibling pair in the proband generation and an offspring all willing to participate in the study. Subsequently, all living relatives from the proband and offspring generations were invited to participate. Spouses of individuals in both generations were also recruited as referents as they share many environmental and lifestyle factors. Newman et al. [28] provides full details of study enrollment and the baseline assessment. Briefly, participants completed sociodemographic and health questionnaires and in-person assessments of physical and cognitive function from 2006 to 2009 (visit 1) and again in 2014–2017 (visit 2). The Institutional Review Boards at all of the Field Centers and the Data Management and Coordinating Center (Washington University, St. Louis) in the USA reviewed and approved this project and the regional Institutional Review Board in Denmark reviewed and approved this project.

The LLFS administered the following neuropsychological tests at both in-person visits: (1) and (2) Logical Memory IA and IIA from the Wechsler Memory Scale–Revised [32] to measure episodic memory using immediate and delayed paragraph recall; (3) and (4) Digit Span Forward & Backward to measure auditory attention and working memory; (5) Digit Symbol Substitution Test (DSST) of the Wechsler Adult Intelligence Scales-Revised [33] to quantify psychomotor processing speed, attention, and working memory; and (6) category fluency (animals) to measure semantic processing [34]. Tests marked as invalid because of poor hearing or vision, environmental distractions, experimenter errors, or other physical limitations were omitted from analyses (n = 202). The longitudinal measurements of each test score were used as the main outcomes. Mini-Mental State Examination [35] was administered in the study and is used to describe the global cognition of the sample, but was not included in the longitudinal analyses due to lack of variability in the study.

Participants’ characteristics at enrollment were summarized using mean and standard deviation, and t-tests were used to compare LLFS family members and their spouses. The goal of our analysis was to test the hypotheses that familial longevity is associated with different neuropsychological test performance and with different rates of longitudinal change, and whether the effect of familial longevity changes with age and generation. We reframed this set of hypotheses as a statistical model selection problem. We initially used Bayesian hierarchical models to describe the individual trajectories of each of the 6 test scores as a function of age at enrollment, follow-up time in years, sex, level of education, field center, an indicator of familial longevity (represented by the spouse variable), a generation indicator (represented by the ind_1935 variable), and the 8 two-way interactions of age at enrollment and follow-up time with sex, education, familial longevity, and generation. The full model that included all main effects and these 8 two-way interaction terms had the following form

y_ij= β₀ × (1 – rep.ind_i) + β_0i × rep.ind_i+ β_age × age.b_i + β_dage × dage_ij + β_sex × sex_i + β_spouse × spouse_i + β_educ × educ_i + β_ind1935 × ind_1935_i + β_{field center} × fc_i + β_sex*age × sex_i × age.b_i + β_sex*dage × sex_i × dage_ij + β_educ*age × educ_i × age.b_i + β_educ*dage × educ_i × dage_ij + β_spouse*age × spouse_i × age.b_i + β_spouse*dage × spouse_i × dage_ij + β_{ind_1935*age} × ind_1935i × age.b_i + β_{ind_1935*dage} × ind_1935i × dage_ij + ε_ij

where y_ij represents the score at time j (j = 1, 2) of participant i and ε_ij is the normally distributed random error with constant variance that was assigned an inverse Gamma prior distribution. The model intercept consists of the fixed effect β₀ and random effects β_0i to account for repeated measurements. We created the variable rep·ind_i to indicate whether participant i had more than 1 assessment (rep·ind_i = 1 for repeated measurements and 0 otherwise). This parametrization assigns a fixed intercept for participants who only had baseline assessment, and random intercepts for those who had 2 assessments. The random intercepts β_0i were assigned Normal prior distribution with mean β₀ and common precision τ, where τ had Gamma prior distribution with both shape and scale parameters equal to 1, which corresponds to assuming an extremely vague prior distribution. We used random intercepts to account for repeated measurements but did not use random slopes in the model to be able to include observations from every individual with at least 1 measurement of a test and prevent bias in the model estimates. All the remaining regression coefficients were assigned Normal prior distributions with mean 0 and precision 0.1. The covariates age·b_i and dage_ij represent the age at enrollment and the follow-up time of participant i, with dage_ij = 0 when j = 1. The variable educ_i was a categorical variable with values 0–17, which approximated years of education. sex_i was a categorical variable with value 1 for male and 0 for female. Three dummy variables were created for each level of the field center variable, using the Boston site as the referent group. The variable spouse_i was coded as 1 to denote a spouse participant and 0 to indicate a long-lived family member. The variable ind_1935_i took on value 1 if a participant was born after 1935 (the younger generation) and value 0 if born on or before 1935 (the older generation). We chose this cutoff based on a previous analysis of the LLFS [36], where the birth year cutoff was used to distinguish the older and younger generations in the LLFS participants. The 8 interaction terms represent effect modification of familial longevity, generation, sex, and education on the age effect at enrollment and on the longitudinal change during the follow-up. With this model formulation, the hypotheses we were interested in testing become hypotheses on specific regression coefficients. Specifically, whether familial longevity is associated with different rates of change of neuropsychological test scores over follow-up is β_spouse*dage ≠ 0, whether familial longevity is associated with different age effects at enrollment is β_spouse*age ≠ 0, and, if these interactions are 0, whether familial longevity is associated with different neuropsychological scores at enrollment is β_spouse ≠ 0.

To test these hypotheses in an unbiased way, we implemented an automated backward search algorithm to select the best model for each neuropsychological test that iteratively drops out the most insignificant interactions from the model. Please refer to the online suppl. material (for all online suppl. material, see www.karger.com/doi/10.1159/000514950) for full details of the algorithm.

In order to assess whether the effect of familial longevity was different in the 2 generations, we added 2 three-way interaction terms β_{ind_1935*spouse*age} × ind_1935_i × spouse_i × age·b_i and β_{ind_1935*spouse*dage} × ind_1935_i × spouse_i × dage_ij to each of the selected models. The added three-way interaction terms were kept in the model if they had a 95% credible interval not including 0, as well as lower order terms of the kept interaction terms. If any of the three-way interaction terms were significant or borderline significant, we further investigated the effects in the younger generation (ind₁₉₃₅ = 1) by running a separate analysis of the study participants born after 1935 using the same model searching algorithm described above.

All analyses were conducted in R3.5 and all Bayesian models were analyzed using the rjags package. The LLFS data used in this article were frozen by June 2018.

Tables 1and2 summarize the demographic characteristics and cognitive test scores of the 5,064 LLFS participants included in this analysis, after excluding 22 participants with missing demographic information. In the older generation (n = 1,959), age at enrollment ranged from 71 to 110 years, with a mean of 88.4 years (SD: 7.6). In comparison to their spouses, older generation participants with familial longevity were older (89.3 vs. 82.6 years, p < 0.001) and a larger proportion were males (46.5 vs. 38.5%, p = 0.01). In unadjusted analyses, LLFS participants with familial longevity from the older generation had significantly lower scores on animal fluency, DSST, and Logical Memory immediate recall at both visits and lower Logical Memory delayed recall at baseline (p < 0.05) compared to spouses. In the younger generation (n = 3,105), age at enrollment ranged from 25 to 79 years, with a mean of 60.0 years (SD: 7.3). LLFS participants with familial longevity in the younger generation were younger (59.8 vs. 60.6 years, p = 0.01) and included a smaller proportion of males (42.4 vs. 48.8%, p = 0.002) than the spouses, at enrollment. In unadjusted comparisons, younger generation LLFS family members had higher scores on DSST and Digit Span Forward and Backward at both visits (at 0.05 significance level) than the spouses.

The 47% of participants from long-lived families who did not have a second in-person assessment were older at enrollment (80.0 vs. 69.2 years, p < 0.0001), were less likely to have a college education (35.7 vs. 54.2%, p < 0.0001), were more likely to have died during the follow-up period (up to April 2018) (61.4 vs. 6.9%, p < 0.0001), and had lower baseline scores on all cognitive tests as shown in online suppl. Tables 1 and 2. Dropout among the spouses exhibited a similar pattern of characteristics as the family members except that there were no differences in education or Digit Span Forward score at baseline. Although dropout selected out the less healthy individuals, this selection was comparable in members of long-lived families and spouses, and the same pattern of differences in neuropsychological test scores between the 2 groups was maintained at visit 2 as shown in online suppl. Figures 1 and 2.

Results of the analysis of each cognitive test score are summarized in Tables 3,4, and5 (these 3 tables show estimates of variables of interest; complete parameter estimates can be found in online suppl. Table 3). DSST was the only test with a borderline significant result of the three-way interaction term between generation indicator, spouse, and follow-up time β_{ind_1935*spouse*dage} = −0.32, 95% CI: −0.67; 0.03), suggesting differential effects of familial longevity in each generation as shown in Table 6 (Table 6 shows estimates of variables of interest; complete parameter estimates can be found in online suppl. Table 4). Figure 1 shows predicted scores for each cognitive test over a continuous age trend by longevity groups, and DSST was the only test that showed different rates of decline (slopes of predicted scores over age). The results of the analysis of each test are summarized below.

Only age at enrollment, follow-up time, education, and sex appeared to have an effect on animal fluency score (Table 3). In the older generation, the animal fluency score was 0.17 points lower for each additional year of age (95% CI: −0.19; −0.15) and 0.06 points lower (95% CI: −0.09; −0.03) for each year of follow-up time. In the younger generation, there were no age effects and follow-up time had a positive effect on the animal fluency score (dage effect = 0.18, 95% CI: 0.11; 0.25), potentially reflecting a learning effect from the first to the second assessment. Higher educational attainment was associated with a higher score; however, the positive effect of education was reduced in older individuals at enrollment (education by age interaction = −0.01, 95% CI: −0.01; −0.005). Males outperformed females, and the difference in scores increased with each additional year of age (sex by age interaction = 0.04, 95% CI: 0.02; 0.05).

Age at enrollment was significantly associated with DSST score in both generations (older generation age effect = −0.68, 95% CI: −0.71; −0.64; younger generation age effect = −0.38, 95% CI: −0.47; −0.29, in female members of long-lived families, Table 3). The negative age effect was higher in spouses compared with members of long-lived families (spouse by age interaction = −0.06, 95% CI: −0.11; −0.01). Education had a positive effect on DSST score that diminished with increasing age at enrollment (education by age effect = −0.01, 95% CI: −0.01; −0.001).

Following the significant three-way interactions between spouse, generation, and follow-up time, an analysis restricted to the younger generation showed that familial longevity decreased the rate of longitudinal decline; spouses appeared to lose an additional 0.13 points for each year in the study compared with members of long-lived families (95% CI: −0.25; −0.01, Table 6). This difference adds up to a 6% additional decline from the baseline average score (51.4) among the spouses compared to LLFS family members over 25 years.

Only age at enrollment, follow-up time, sex, and education appeared to have cross-sectional effects on the Digit Span tests (Table 4). In the older generation, older age at enrollment and longer follow-up time were associated with lower Digit Span Forward and Digit Span Backward scores (Forward age effect = −0.03, 95% CI: −0.03; −0.02; Forward dage effect = −0.12, 95% CI: −0.13; −0.11; Backward age effect = −0.03, 95% CI: −0.04; −0.02; Backward dage effect = −0.04, 95% CI: −0.05; −0.03). In contrast, only the Digit Span Forward score was associated with age at enrollment and follow-up time in the younger generation and the effects were opposing (age effect = 0.02, 95% CI: 0.003; 0.04 and dage effect = −0.07, 95% CI: −0.09; −0.04). Males scored higher than females on Digit Span Forward and education was positively correlated with both tests with a similar magnitude of effects.

The effect of age on both immediate and delayed recall conditions was opposing for the 2 generations and was modified by education and sex (Table 5). In the older generation, older age at enrollment was associated with lower immediate and delayed recall, but was associated with higher scores in the younger generation. Additional years in the study were associated with higher scores in both generations with similar magnitude of effects between recall conditions, most likely due to practice effects. Male sex was associated with lower scores in both tests, although the negative effect was reduced with older age at enrollment (immediate recall sex by age interaction = 0.01, 95% CI: 0.002; 0.03; delayed recall sex by age interaction = 0.03, 95% CI: 0.02; 0.04). Higher educational attainment was associated with higher scores. However, the positive effect of education was reduced with older age in both tests (education by age interaction = −0.003, 95% CI: −0.01; −0.001). In both recall conditions, spouses had lower scores compared with members of long-lived families (immediate recall spouse effect = −0.38, 95% CI: −0.63; −0.13; delayed recall spouse effect = −0.39, 95% CI: −0.65; −0.13), but we did not detect any difference in the rate of longitudinal change.

This study used Bayesian hierarchical modeling of longitudinal data to describe changes in 6 neuropsychological tests measured in two generations of LLFS participants. We found that participants with familial longevity outperformed their spouses on the DSST and the immediate and delayed recall conditions of the Logical Memory test at enrollment indicating better cognitive function in the domains of psychomotor processing speed and episodic memory and retrieval, respectively. The LLFS family members also had a smaller negative age effect on the DSST of 0.06 points per year than their spouses, which, for example, would result in a 4.2 points higher score for a 70-year-old member of a long-lived family compared to a spouse of the same age. Additionally, in the younger generation, family members declined more slowly than spouses on the DSST by 0.13 points per year, on average increasing the performance advantage seen of 5.9 points at visit 1 to 6.9 points at visit 2.

Our results are consistent with a previous analysis by Newman and colleagues [26], which found better performance among LLFS family members on the DSST in comparison to their spouses (analyzed in the offspring generation only) and to participants of the CHS (both generations). The current study extends the findings of a baseline difference in DSST performance between family members and spouses to the older generation and adds findings of a baseline performance advantage for family members on the immediate and delayed recall conditions of Logical Memory as this test was not analyzed in the previous study. Our results are, however, inconsistent with a previous analysis of cognitive function among a subset of individuals in the offspring generation in LLFS which found better performance among offspring on the Digit Span tests and an alternate category fluency test (i.e., vegetables) and no difference on animal fluency or Logical Memory [30]. However, the previous study used a restricted group of offspring (n = 305) consisting of only the offspring of the proband and did not include the DSST in the analysis and thus, the findings of that analysis may be specific to the offspring of the oldest LLFS siblings (i.e., nonagenarians and centenarians). In this work, we expanded the analysis to include all study participants resulting in more statistical power and results that can be generalized to family members within two generations of long-lived families. The Leiden Longevity Study, which studies the offspring of nonagenarian siblings, noted better performance on immediate and delayed recall of pictured items but no difference on the DSST when comparing offspring with their spouses [37]. Our findings of better baseline performance on Logical Memory among individuals with familial longevity is consistent with their study and suggests that familial longevity may be associated with lifelong differences in episodic memory. Whereas the current study also found a performance advantage among family members on the DSST, the lack of an association in the Leiden study may be due to the much smaller sample size (offspring n = 250 compared to n = 2,283 in the current study) or differences in the definition of familial longevity (sibling pairs with longevity vs. survival probability calculations of the entire proband generation).

Familial longevity and sporadic longevity (i.e., individuals living to extreme ages) are distinct models of longevity and may occur due to different biological, genetic, and environmental contributors. Yet, the current findings are also consistent with some studies of parental, rather than familial, longevity. For example, centenarian offspring have been shown to have a reduced baseline prevalence and incidence of cognitive impairment compared with cohort-matched referent groups without familial longevity [27]. In contrast, the Health and Retirement Study (HRS) noted a slower longitudinal decline in global cognitive function but no difference at baseline for offspring of octogenarian and nonagenarian parents [38]. The lack of findings in the HRS may be due to the less stringent definition of parental longevity (i.e., survival probability of reaching age 80 vs. 100 years). Regarding more specific cognitive functions, the Framingham Heart Study found better attention at baseline among individuals with a parent who lived to age 85 years or older but did not find better episodic memory recall, as in our current study [26]. They also found that these offspring had slower declines in attention and executive function over longitudinal follow-up; however, they used different cognitive tests than the current study. Our findings, particularly for episodic memory and processing speed, suggest that clustering of longevity within families (e.g., multiple siblings within a family living to age 90+ years) may be a more robust phenotype for studying healthy cognitive aging than studies of individuals with octogenarian or nonagenarian parents.

It is interesting that our analysis could detect significantly different rates of change only for the DSST that were restricted to the younger generation. This test is regarded as a highly sensitive test for detecting cognitive change, particularly at higher levels of cognitive functioning [39, 40]. Notably, the DSST is a multifactorial test, meaning that it can be affected by deficits in many cognitive processes (e.g., attention, processing speed, episodic memory, and visuoconstruction) as well as deficits in noncognitive factors (e.g., motor speed, vision, and motivation). The use of a multifactorial test may be beneficial for capturing the heterogeneous changes associated with aging. It must also be noted that the DSST and similar coding tests appear to be better predictors of death among older adults over follow-up than other cognitive tests [41-43] and chronological age [44]. Since the LLFS cohort is predisposed to longevity, cognitive advantages related to performance on the DSST may be enriched among these individuals. Additionally, the DSST is known to have higher heritability across ages than other cognitive tests [24]. As the LLFS is a family-based study, this suggests that it may have additional power to identify genetic and environmental contributions associated with preservation of cognitive function, particularly for processing speed.

Changes over time on all of the other cognitive tests were similar between LLFS family members and spouses. In both generations, declines over time were seen on Digit Span Forward, whereas performance on Logical Memory immediate and delayed recall improved, most likely due to practice effects. Some tests showed different patterns of change between the two generations; animal fluency and Digit Span Backward performance declined in the older generation, whereas in the younger generation, animal fluency improved and Digit Span Backward remained stable. Although slight age-related declines are generally seen across cognitive domains with the exception of vocabulary [11, 16], it must be noted that the younger generation participants are just beginning to enter the decade at which greater age-related changes in cognitive function are observed [45]. A lack of findings may also be due to prior test experience, which has been shown to have a greater effect on tests of memory and reasoning and which may mask longitudinal changes [10]. In fact, the findings of an association of better Logical Memory scores with a greater number of years in the study supports this notion.

In a broader context, this study reinforces the importance of studying a variety of cognitive domains when assessing cognitive aging. In the LLFS, we now have evidence that individuals from long-lived families may have cognitive advantages for some cognitive processes that persist with aging. Our results also suggest that the association of familial longevity with slower cognitive decline may be restricted to only certain domains of cognitive function. If this is indeed the case, the LLFS cohort will allow for the study of factors that are related to better cognitive function and separately, factors that lead to slower cognitive decline. However, there is a caveat: in order to determine which cognitive functions are innately better among long-lived family members versus which show slower change over time we need to begin studying individuals predisposed to longevity much earlier in the life course since longitudinal studies have shown that cognitive decline beings as early as one’s 40s [11]. Related to this, findings from the Lothian Birth Cohort 1936 suggest that higher intelligence in childhood is associated with better cognitive ability, in one’s 70s [46]. However, they did not find childhood intelligence to be related to rate of cognitive decline. Therefore, identification of a change point or onset of cognitive decline may allow us to determine whether the cognitive advantages in our cohort are due to higher premorbid function and/or slower decline. Furthermore, future work will need to parse out the specific cognitive processes that are captured by the DSST and that appear to be more resistant to aging in the younger generation of the LLFS.

This study has several strengths. The cognitive battery in the LLFS consists of in-person assessments covering most cognitive domains. Additionally, whereas some studies of individuals with familial longevity have only looked at individuals age 65 years and older, the participants in our sample cover a wide age range from 25 to 110 years. Finally, there is a growing need for more sophisticated methods for analyzing longitudinal neuropsychological data [47]. As such, we used a statistical analysis that allowed us to test hypotheses about the effect of familial longevity in an unbiased and reproducible way. Although the Bayesian modeling technique we used is computationally challenging, it enabled us to incorporate all information from the data, including data from participants who underwent only 1 visit. In the model we used, the data of participants with only 1 visit contributed to the estimation of the cross-sectional effect of age, while the data of participants with 2 measurements contributed to the estimation of the longitudinal age effect. The challenge of Bayesian modeling, however, is the lack of well-established criteria for model selection. We used a novel approach to Bayesian model selection that is more robust to false positives.

There are also limitations of this study that need to be acknowledged. First, the spouses of the LLFS family members may not be an appropriate referent group, as individuals tend to find partners who share sociodemographic, lifestyle habits, and even genetic characteristics [48, 49]. Additionally, at an average age of 83 years at enrollment and 87 years at visit 2, spouses in the older generation are a selected group of individuals who have surpassed average life expectancy. Thus, the similarity of spouse pairs and survival bias of the spouses in the older generation may have biased our results toward null findings. Future studies should compare cognitive performance of LLFS family members with individuals who are representative of the general population. Second, there was substantial dropout in the older generation (71% of family members and 62% of spouses) and thus a healthy survivor bias in the longitudinal results. In the younger generation, a smaller proportion of spouses than family members completed both in-person assessments (60.2 vs. 69.5%, respectively), which may have led to a selection of healthier family members. One should also note that not all members of long-lived families go on to reach exceptional ages themselves. We are more likely capturing the subset of individuals who have a greater predisposition to longevity and, therefore, remain in the study and have better cognition. It is clear that participants who dropped out in both generations were older and had lower cognitive test scores, but this was true for both family members and spouses, so we do not expect the dropout to affect the results. Third, our cognitive battery does not assess all cognitive processes with notable exceptions including visuospatial abilities and visual memory. Additionally, other cognitive tests that are more sensitive to age-related and neurodegenerative changes may have revealed other cognitive advantages associated with familial longevity beyond processing speed and episodic memory. Last, the availability of only up to 2 time points of cognitive data restricts our analysis to linear changes in cognitive function as opposed to nonlinear trajectories and does not allow us to rule out that some changes in cognition were due to regression to the mean or that some changes are not able to be accurately captured due to practice effects.

Families with siblings reaching their 80s and 90s appear not only to have better cognitive performance than spouse controls, but they also demonstrate slower decline on a multidimensional measure of psychomotor processing speed. The families in the LLFS may, therefore, be an important resource for geroscientists studying the basic biological mechanisms of aging and more specifically, resilience to cognitive aging. Moreover, family studies, such as the LLFS, provide extra power for examining the genetic factors associated with preservation of cognitive function.

The study subjects gave their written informed consent, and the study protocols were approved by the Institutional Review Boards at all of the Field Centers and the Data Management and Coordinating Center (Washington University, St. Louis) in the USA and the regional Institutional Review Board in Denmark.

The authors have no conflicts of interest to declare.

Supported by the National Institute on Aging (K01AG057798 to S.L.A., 5U19AG063893 5U01AG023749 to S.C., 5U01AG023755 to T.T.P., 5U01AG023712, 5U01AG023744, and 5U01AG023746); the Boston University School of Medicine Department of Medicine Career Investment Award to S.L.A.; and the Marty and Paulette Samowitz Foundation to T.T.P. Portions of these findings were presented as a poster at the 2020 International Neuropsychological Society Annual Meeting, Denver, CO, USA.

S.L.A. and M.D.: statistical analysis and drafted the manuscript; P.S.: designed the study; S.L.A., M.D., S.C., N.S., A.L.R., T.T.P., and P.S.: interpretation of the results. All authors edited the manuscript.

S.L. Andersen and M. Du shared equally in their contribution to this work.