The magnitude and direction of the relationship between risk factor and cognition depends on age: a pooled analysis of 5 community-based studies

Late-life cognitive impairment and dementia (CD) are devastating. The evolution of these conditions begins long before clinical onset so early prevention interventions are needed [1]. Because of their potential for modification and the availability of data from life-course epidemiologic studies, control of cardiovascular risk factors (CVRF) has been targeted as a class of risk factors that could be intervened on to prevent or slow the progression of CD. However, observational and trial evidence of the magnitude of the association between CVRF and CD is mixed, and it is still unclear which CVRF during which age period would be the most effective approach to reduce the burden of CD [2]. This gap has slowed the development of effective interventions on CVRF aimed to reduce the burden of CD.

Evidence of the mixed findings is in the variation in the slopes, risk ratios or significance of the relationship between a CVRF (i.e., blood pressure) and CD outcomes. Reports from observational studies range from an expected higher CVRF level increasing the risk for CD, to a null association, to an unexpected lower CVRF level increasing the risk for CD [3]. The few clinical trials that test the effects on cognitive impairment of intervening on a single cardiovascular risk factor have largely been negative [4‐6] with the exception of SPRINT MIND [7].

Reconciling these mixed findings is hampered because they are based on single studies of different aged cohorts (i.e. ≤ 55; ≥ 55, ≥ 65 or ≥ 75 years. of age), different study designs (i.e. cross-sectional, case–control or longitudinal), different target populations (i.e. clinical volunteers, patient groups, or population-based), different outcome measures (i.e. cognitive testing, clinical diagnosis, or algorithmic classification) and follow-up time as well as different statistical models to test for associations. Across cohorts, there may be a different balance of confounding variables, which could bias not only the internal validity of the study, but also the validity of cross-cohort comparisons. It is also a challenge to study early risk factors because CD is, by definition, a condition of late-life with initially slow progression [8].

Age of an individual or a study cohort is a core feature that can account for the inconsistencies in the literature. Here, we are interested in the extent to which, and the pattern of how, age modifies the magnitude and direction of the association between CD and CVRF. This question differs from those that include age as an independent correlate, as a confounder, or as a scale to show longitudinal trajectories of a variable. Testing for this age interaction is similar to studies presenting age-stratified analysis to account for possible differences in the association of a CVRF and cognitive outcome by age strata [2, 3]. However, such age-stratified analyses often include a wide age range and assume a constant association between the risk factor (RF) and outcome over the age interval. Further, the studies of mid-life CVRF to late-life cognition can have up to a 15–20 years gap between measures, so there are few data giving insight into the evolution of change in the association. Understanding this evolution can help design future prevention and precision medicine approaches and also contribute to our interpretation of observational studies.

Here we aim to test the assumption of a constant association of RF to CD over a wide age range. We calculated 1 year slopes of these associations for five CVRF and modeled the trajectory of these 1 year age slopes from mid-to late-life. These 1 year slopes are derived from pooled individual-level data from five mid- and late-life community-based cohorts that have measured the same cognitive test and set of CVRFs. We modelled the trajectory of these 1 year slopes to investigate whether there is a negative association between CVRF and CD throughout mid- to late life or whether these associations change over time.

The percent women was generally similar across cohorts (range 42.1–46.8%). At baseline (Table 1), ARIC and CARDIA had the youngest participants and AGES-RS the oldest. Trends in CVRF and DSST across cohorts generally tracked with the relative ages of the cohorts. The mean and standard deviation of harmonized and unharmonized variables are shown in Online Resource 13. The mean of the observed DSST-CVRF 1 year slope per age bin, the modelled trajectory of the 1 year slopes, the smoothed trajectory of the 1 year slope, and 95% CI of the trajectory are shown in Fig. 1a–e. The predicted and smoothed trajectories by study and by CVRF are shown in Online Resource 14.

For DBP, the linear-quadratic model (Fig. 1a) had the lowest BIC, with agecut = 67. Over the age span, there was considerable variation in the 1 year slopes. For younger ages (less than 55 y) and older ages (≥ 67 year) 1 year slopes were mostly negative (i.e., < 0), suggesting higher DBP is associated with lower DSST scores. Between 55 and 70 year, 1 year slopes attenuated and became positive close to zero (p = 5.3E−3). Although still positive at age 67, after this, slopes start a downward trend and become increasingly negative until approximately 80 years of age, when again the 1 year slopes increasingly attenuate to zero (Tslope2sq = 0.00072 (0.00045,0.00099; p = 2.4E−7). This translates into an estimated 0.013 higher DSST score at age 60 years; a 0.014 lower DSST score at age 70 years; and a 0.079 lower DSST score at age 80 years for each increment in 1 mmHg DBP, controlling for covariates.

Similar to DBP, the SBP (Fig. 1b) linear-quadratic model had the lowest BIC, with agecut = 67 years. Individual 1 year SBP-DSST-slopes were negative across the age span but became less negative as age reached 67 years and then became more negative as age reached the relative minimum (nadir), of 75.4 years. After age 75 years, the 1 year slopes attenuated with age. Together these changes reflect a U-shaped relationship from age 67 to 88+ years old (T-slope2sq = 0.00029 (0.00014,0.00044; p = 1.3E−4). Based on this model, with each increment in 1 mmHg SBP, adjusting for covariates, there is an estimated 0.021 lower DSST score at age 60 yo; a 0.027 lower DSST at age 70 yo; and at age 80 yo there is a 0.029 lower DSST score.

The BMI trajectory was best modeled (lowest BIC) by the linear–linear model (Fig. 1c) analysis, with agecut = 70 years. For age ≤ 70, for each unit increase in the 1 year slopes of DSST-BMI, there was an average 0.12 lower DSST; however, the trajectory is not significant (p = 0.5) and approximately constant. After age 70 years, 1 year-slopes become increasingly less negative with age (p = 1.5E-3), attenuating to zero and becoming positive slopes around age 80 years, i.e., a higher BMI was associated with a higher DSST score after 80 years of age. For each increment of 1 BMI unit, this translates to an estimated 0.11 lower DSST score at age 60 years; a 0.10 lower DSST score at age 70 years; and a 0.009 lower DSST score at age 80 years.

The best fitting model for total cholesterol was the linear–linear model (Fig. 1d) and agecut = 69. Overall, 1 year slopes were constant and close to zero for age < 69 years (p = 0.29). After age 69 years, the magnitude of the 1 year-slopes of the relationship between DSST and cholesterol tended towards more negative such that a higher cholesterol was associated with a lower DSST score (p = 2.2E-3). This translates to an estimated 0.0016 higher DSST score with each increment in 1 mg/dL of cholesterol at age 60 years; a 0.0029 higher DSST score at age 70 years; and 0.0056 lower DSST score at age 80 years.

The linear–linear model for fasting glucose (Fig. 1e), had the lowest BIC, with agecut = 75 years. Overall, 1 year slopes for DSST-fasting glucose associations were negative, meaning that for a unit increase in fasting glucose, there was a lower DSST score. The linear trends before (p = 0.07) and after (p = 0.057) age 75 were relatively constant but attenuate to zero. For age ≤ 75 years, the expected 1 year-slopes were marginally lower by 0.00035 for each increase in 1-unit of fasting glucose. After age 75 years, the expected 1 year-slopes was 0.0015 points higher with each 1-unit increment of age, with a drift of the 1 year slopes attenuating towards zero. This translates into an estimated 0.021 lower DSST score at age 60 years; a 0.024 lower DSST score at age 70 years; and a 0.019 lower DSST score at age 80 years for each increment in 1 unit of fasting glucose level.

In 2018, the National Academy of Science summit on prevention of cognitive decline [Preventing Cognitive Decline and Dementia: A Way Forward, National Academies Press, http://​www.​nap.​edu; and https://​www.​nia.​nih.​gov/​research/​administration/​recommendations-nih-ad-research-summit-2018] concluded the evidence for control of CVRF to reduce the risk for CD is not robust or consistent enough to support prevention trials or for public health messaging. The design of effective pharmacological or behavioral approaches continues to be priority and to move forward, more insight is needed on factors underlying the incongruent results.

Although this analysis does not address the specific reasons for the direction and magnitude of changes in the slope trajectories, several contributing factors to the age-related loss of the CVRF explanatory or predictive power can be proposed. First, as a result of the increase in multi-morbidity with aging, there may be a proportional reduction in how much any one risk factor can explain CD. An oft suggested explanation is ‘reverse causality,’ where one hypothesizes that X ‘causes’ Y, but in reality, Y actually ‘causes X.’ For instance, we find high blood pressure is associated with better, not worse, cognitive decline or we find lower weight and cholesterol levels associate with an increased, not decreased, risk for dementia [24, 25]. This reverse causation could reflect underlying biologic, behavioral or methodologic issues [26, 27]. In the case of dementia, it is reasonable to assume behavioral changes, such as loss of appetite or more sedentary behavior, may affect biologic changes resulting in CVRF levels that no longer reflect the past cumulative exposure to the risk factor. Biological changes can result in reverse causality if there is a shift in the balance between the central and peripheral regulation of CVRF, whereby the pathology in the brain regulates the risk factor, and not vice versa. For instance, neurodegenerative pathology in the hypothalamus may affect regulation of glucose, insulin or body weight [28], leading to prodromal changes such as more rapid weight loss and lowering of blood pressure and cholesterol levels as compared to those who maintain cognitive function [24, 29]. Although data are limited on the empirical significance biological changes, epidemiologic and statistical studies suggest failure to account for possible reverse causation can lead to biased conclusions [30].

The variability in the BP trajectories, particularly the DBP during the 70 to 80 years old band, is striking and significant (the quadratic term in the piecewise modeling of the dip is statistically significant). Although standardized methods were used and we harmonized the BP measures across cohorts, blood pressure measures per se are very variable, partly reflecting measurement issues for SBP and DBP, with DBP having added sources of reading error when sphygmomanometers [31] are used. Additionally, at older ages there are many factors that can influence participation and functional performance. With age, SBP tends to increase, while DBP may decrease due to vascular disease such as increased arterial stiffness [32]. Further, compared to the other risk factors we examined, high BP is a more direct measure of mortality risk, so selection bias may be relatively stronger for those with unfavorable BP levels [33, 34] compared so the other CVRF. As result of all these factors, the sample composition in the bracket between 70 and 80 of age could be very heterogeneous across and within studies and this may be reflected in the variability of the trajectory of 1 year slopes.

Our novel approach uses the trajectory of slopes based on pooled population-based cohorts to frame issues that are important to designing future research. [35] Our study points to candidate CVRF, estimates age intervals when the impact of an intervention may be more or least effective, and provides some insight into how long an intervention, how large a sample size, or how big of a treatment effect is needed to achieve a detectable effect. For example, if the CVRF-Cognition association is strongly negative during a particular age interval, then an intervention either before or during the age interval may be more likely to positively change the outcome than during an age period where there is no association. Our study also suggests the effect sizes of any trial are likely to be small so clinical trials conducted at later ages will need large sample sizes or target vulnerable sub-groups and will need to deliver an intervention that creates large differences between treatment arms, such as SPRINT MIND, where the difference in SBP in the treatment versus control arms was > 14 mmHG [36]. Further, the attenuation of the relationships suggests the contributions to poor cognitive function of other risk factors is increasing; this multi-morbidity should be accounted for in patient selection and should be reflected in the extrapolation of the results to more complex populations. Finally, this research can also inform the comparison and integration of findings from the increasing number of observational studies reporting on the CVRF-CD associations. Importantly, if the association varies depending on the age of the cohort, it is reasonable to further investigate any methodologic reasons for the variation before abandoning a candidate RF or before planning an intervention for the RF.

Our analysis has several strengths. By including studies with similar study designs, standardizing the interval between the CVRF and DSST, no imputation of cognitive scores, and applying similar analytical approaches to all the studies, we have minimized these sources of variability present in cross-study comparisons. We also allowed for non-linear 1 year slope trajectories and used BIC criteria to select the best fitting models so we could better capture the patterns of change in slopes over time.

However, several aspects of our study should be noted when interpreting the results. The trajectories are based on cross-sectional binned 1 year slopes of CVRF–DSST associations so the analyses are subject to the same biases as other cross-sectional studies of ‘change by age’—most importantly the bias caused by selective dropout by age and health condition. As a result, there is the possibility that the sample is getting ‘healthier’ with age and this may explain some of the attenuated associations of CVRF to DSST. We used a single cognitive test with a normal distribution of test scores reflecting psychomotor speed and executive function; tests with different properties may give different results. It is also possible that over the 30 years spanned by these 5 cohorts there were secular trends in the CVRF that may affect the magnitude of the associations, but because the cohorts did not perform cognitive testing at all exams, we could not test this [37]. However, a recent article [38] reporting on a NHANES analysis of serial cross-sectional exams from 1999 to 2018 showed there were only ‘sub-optimal’ improvements over time in CV health. Finally, we incorporated relatively few confounding variables in our models, including medications to lower the RF, but our analysis showed including more confounders the model fit did not improve. Never-the-less, we cannot exclude the presence of residual confounding.

In conclusion, understanding the methodologic and demographic differences among observational studies will aid in the interpretation of CVRF—CD associations and promote more targeted designs of clinical trials to prevent cognitive loss in late-life.