The pace of memory decline in older adults without a neurocognitive disorder: Results from 22-years of follow-up in a nationally representative sample

Abstract

Introduction:

The pace of cognitive change is one of the major questions in cognitive aging. The Children of the Depression Age (CODA) cohort of the Health and Retirement Study (HRS) is uniquely suited to study cognitive aging because it has a long follow-up (22 years) and a narrow age range at baseline (67–74 years), and presents a unique opportunity to study this topic.

Methods:

We examined delayed recall data over the 22 years of follow-up in a nationally representative sample of the United States (HRS-CODA; N = 2,295 at baseline and N = 263 at the final follow-up wave), examining results for the entire sample and omitting participants with self-reported dementia. Data were analyzed using latent growth curve models, adjusting for baseline age, sex, years of education, and race/ethnicity.

Results:

Respondents were predominantly female (62%), White (86%), and 71 years old on average at baseline. Our results suggest the pace of normative (defined as the absence of a dementia diagnosis over the follow-up period) memory decline is about −0.05 standard deviations per year (SD/y), but is better characterized by age specific estimates of −0.04 SD/y, −0.10 SD/y, and −0.15 SD/y for an individual who was 75, 85, and 95, respectively.

Discussion:

Memory decline, in the absence of a recognized dementia and without a confounding of baseline age differences and longitudinal age changes, would be present but almost imperceptible to an individual in their eighth decade, but noticeable in their ninth and quite impairing in their tenth decade. Future research is needed to examine other cognitive domains and with more robust measures.

Introduction

On average, cognitive ability declines later in life [1–3], although there is substantial variability in the amount of decline by individual and by cognitive domain [4, 5]. Neurocognitive disorders, such as Alzheimer’s Disease and Alzheimer’s disease related dementias (AD/ADRD), as well as mild cognitive impairment or cognitive impairment no dementia (MCI/CIND), are defined by accelerated cognitive decline [6–9]. However, the pace of normative cognitive aging, defined as cognitive aging in the absence of neurocognitive disorders, remains understudied [1, 10]. This lack of knowledge on the pace of normative cognitive aging presents a major issue in cognition research, as there is no established reference to serve as a point of comparison for research on pathological cognitive aging in clinical populations.

Previous studies on normative cognitive aging, from both longitudinal and cross-sectional studies, have suggested a pace of about −0.01 to −0.15 standard deviations per year (SD/y [3, 11–14]). However, there are limitations to these estimates. For example, Salthouse and Ferrer-Caja (2003) drew estimates from cross-sectional samples. The effect of age is difficult to interpret given the cross-sectional age differences are confounded by cohort effects, secular changes, and differential survival effects [1]. Hayden et al. (2011) examined data from a sample of clergy members for up to 15 years, concluding that cognitive aging is best studied over decades. However, their sample included participants ranging in age from 56 to 85+ at baseline, and this age heterogeneity makes interpreting the results difficult as individuals who are older in this cohort may not be representative of the younger participants, as not all the younger participants may survive until their mid to late 80s (i.e., differential survival). Age was used as the time metric in Hayden et al. (2011), which confounds cross-sectional age differences and longitudinal age trends [15]. Cullum et al. (2000) drew from a convenience sample of community-dwelling older adults, but their follow-up period was between 3.5 to 5 years in total, which may be too short to provide rigorous estimates of decline [3]. There are also methodological issues such as floor and ceiling effects, non-linear scaling properties of cognitive variables which can influence the estimation of cognitive aging trends, which are not accounted for in some of these studies [16–18].

The Children of the Depression (CODA) cohort of the Health and Retirement Study (HRS) is a unique opportunity to capitalize on a cohort with age homogeneity at baseline and a long follow-up period to address the issues of studying normative cognitive aging. HRS-CODA began in 1998 and currently has 22 years of follow-up data available. Second, HRS-CODA is a nationally representative community sample. Third, with age homogeneity at baseline, HRS-CODA cohort simplifies the interpretation of aging effects by minimizing the confounding of cross-sectional age differences and longitudinal aging effects. Given these strengths, the HRS-CODA cohort is ideal for studying normative cognitive aging in older adults. This study aims to observe changes in memory over 22 years in the HRS-CODA cohort and provide an estimate of normative memory decline.

Methods

Participants

Of N = 2,464 participants in the HRS-CODA cohort, n = 6 were dropped for having a baseline age outside of the age band of 67–74, and n = 163 were dropped for having a survey weight of zero. Survey weights of zero meant the participant was ineligible for the study, which is typically because an ineligible spouse of an eligible participant was also interviewed (see the Procedures section for more details). This brought the final analytic sample to N = 2,295. The sample was predominantly female (62%), White (86%), and non-Hispanic (94%), and were 71 years old at baseline with 12 years of education on average. Table S1 shows the sample characteristics. In 2022, n = 263 participants, or 11% of the baseline sample, remained in the cohort. Informed consent was obtained by the HRS for study participation.

Measures

The primary outcome variable was a measure of delayed recall of 10 nouns, with a score range of 0 to 10 [19]. Although the range of possible scores on the delayed recall task was 0 to 10, few participants scored above 8 overall. We therefore discretized the upper end of the delayed recall measure, resulting in a range of 0 to 7, and treated delayed recall as a categorical outcome variable in the analyses due to the large floor effect shown in Figure 1. By treating delayed recall as categorical, we acknowledge that the differences between a score of 0 and 1 (i.e., the floor effect) has a different implication for memory ability than the difference between other scores. These decisions are further explained in Supplement 1.

Fig. 1. Delayed word recall response categories and number of words recalled, Health and Retirement Study Children of the Depression Age (HRS-CODA) sample, 1998 wave (N = 2,295).

Note: Unweighted proportions are shown in Figure 1.

An additional study variable asked participants if a doctor has ever told them they had a memory-related disease (1998 – 2008) or if a doctor has ever told them they had dementia or Alzheimer’s disease (2010 – 2020).

Procedure

The HRS began in 1992, consisting of participants aged 51–61 and their spouses of any age living in the United States. The HRS merged with the Asset and Health Dynamics among the Oldest Old (AHEAD) in 1998, consisting of participants 70 or older, and began enrollment for the CODA cohort. The HRS-CODA cohort was restricted to participants born between 1924 and 1930. Since 1998, the HRS has gone on to use a steady-state design, enrolling a new cohort every 6 years. HRS assessments take place every two years [20, 21].

Sampling in the HRS is done at the household level, where a primary respondent is randomly selected among all age-eligible household members, and their spouse or partner is also interviewed, regardless of their eligibility. The HRS oversamples African-American and Hispanic households, and sampling weights are derived to account for differing probability of selection and non-response in each wave, accounting for geographic stratification and clustering [20, 22].

To assess word recall, participants were first read a list of 10 nouns by the interviewers and repeated them immediately (immediate recall). After about a five-minute delay, participants are asked to repeat the 10 nouns again (delayed recall). Four different word lists were used, and participants were randomly assigned one of the lists at their first assessment. At follow-up assessments, the participants were assigned one of the three remaining word lists, which continued iteratively until all four word lists had been used in subsequent assessments. If two participants in the same household were eligible for the HRS, they were not presented the same word list.

Participants were randomly assigned to have their baseline assessment be in-person or via telephone, but participants were free to switch to another modality based on their preference. In later waves, participants were also able to complete the assessment online. However, less than 1% of HRS-CODA participants completed an online assessment during follow-up, so we combined the few online assessments with the telephone assessments for the purpose of this study. See Table S2 for mode of assessment by time point.

Analyses

We used a latent growth curve model as the primary analytical model for this study. We accounted for mode of assessment as a time-varying covariate given previous research identifying differential item functioning effects of mode of assessment on cognition in HRS-CODA [23, 24]. We adjusted for sex, race/ethnicity, years of education, and age at baseline, and tested for practice and retest effects with a third latent variable [18]. Time was coded in decades for the analysis. See Figure S1 for a depiction of the final model and Table S3 for the design matrix. We incorporated the survey weights developed by the HRS in 1998 into the analyses, and accounted for the stratified survey design. We ran two versions of this model, one where we included the entire HRS-CODA sample, and one where we removed participants who self-reported a memory problem, dementia, or Alzheimer’s disease diagnosis over the entire follow-up period. No other exclusion criteria were used when creating the sample for these models.

Delayed recall was treated as a categorical variable in the latent growth curve model. The advantage of this approach is that there is no assumption of uniformity in the difference between response categories (i.e., the difference between 0 words correct and 1 word correct is assumed to imply a different magnitude of memory ability difference relative to the difference between 5 words correct and 6 words correct). The analysis model is an ordered logit, and the latent response variable is the outcome for the growth curve model. The latent response variable is assumed to be normally distributed, and can be standardized which allows for examining standardized change throughout follow-up, even though the outcome variable itself is categorical.

To determine if the model showed good fit to the data, we evaluated the χ² test, Tucker-Lewis Index (TLI), root mean square error of approximation (RMSEA), and standardized root mean square residual (SRMR). A non-significant χ² test indicates best fit to the data, however, the χ² test is highly sensitive to model misfit and a significant test alone is not enough to indicate poor model fit [25] if other fit indices show good fit. Ideal TLI values are < .95, and ideal RMSEA and SRMR values are <.05 and <.08, respectively [26–29]. Missing data were handled using full-information maximum likelihood estimation. Analyses were conducted using R version 4.4.3 (R Core Team, Vienna, Austria) and Mplus version 8.10 (Muthén & Muthén, Los Angeles, CA).

Results

Latent growth curve model results

We fit several growth curve models with linear and quadratic slopes to the data, which are explained in Supplement 2. The best fitting model to the data included a linear and quadratic slope, but no practice or retest effect. Baseline age was centered at 70 years old, and all other continuous variables were centered at the mean for interpretation. The model for the full HRS-CODA cohort showed excellent fit to the data, χ² (303) = 422.28, p < .001, TLI = .98, RMSEA = .01, SRMR = .04, as did the model where we omitted participants with a self-reported memory problem, dementia, or Alzheimer’s disease diagnosis (N = 477 omitted), χ² (303) = 375.35, p = .003, TLI = .98, RMSEA = .01, SRMR = .04. Coefficient estimates are available in Tables S5 and S6, and the models are shown in Figure 2.

Fig. 2. Cognitive trajectory in the HRS-CODA cohort (N = 2,295).

Note: Figure 2 shows the results of the primary analyses, suggesting cognitive aging is faster at older ages (e.g., 90s) compared to younger ages (e.g., 70s). In panels A and B, the black line refers to the sample where we removed any participant with self-reported dementia or memory problems at any point in their up-to 22 year follow-up, and the gray line is the entire HRS-CODA sample. The dashed lines extend the current trajectory past the 22 years of follow-up to 25 years of follow-up. Panels C and D provide the paces of cognitive aging calculated from the quadratic (panel C) and linear (panel D) models for the full cohort and the sample with no self-reported dementia or memory problems.

Calculating pace of normative memory change

Using the quadratic model, the instantaneous slope at a given point along the trajectory can be determined by taking the derivative of the linear combination of the slope and quadratic effects. With age centered at age 70, we calculated the instantaneous slope at 5, 15, and 25 years after baseline to be estimates of normative memory decline for an individual who was 75, 85, or 95, respectively. In addition to the quadratic model, we provide a singular pace of memory decline from the linear model. However, as shown in Figure 2A and 2B, the linear model underestimates the pace of memory change compared to the quadratic model. Thus, while the linear model provides an overall result that could be useful when examining cohorts with age heterogeneity, the better estimate of memory decline comes from the quadratic model where the pace of cognitive change varies by age. The paces of normative memory change in both samples are shown in Figure 2C and 2D. The results suggested paces of normative decline of −0.04 standard deviations per year (SD/y) for an individual who was 75, −0.10 SD/y at 85, and −0.15 SD/y for someone who was 95 in the full sample based on the quadratic model, and −0.05 SD/y as an overall estimate from the linear model. These standard deviations refer to the underlying latent response variable from the latent growth curve model. Using a 0.50 SDs as a benchmark for a clinically meaningful difference [30], at −0.04 SD/y it would take 12.5 years to reach a change of 0.50 SDs. At −0.10 SD/y it would take 5 years, and at −0.15 SD/y, it would take 3.33 years. Thus, normative change happens at a slow pace in the 70s, but rapidly accelerates to noticeable change every 5 years in the 80s, and about every 3 years in the 90s. Using the linear model, noticable change would occur each decade.

Effect of covariates on cognition

At baseline, women, non-Hispanic White participants, and participants with more years of education had higher delayed recall ability whereas older participants had lower delayed recall ability. To test the effect of covariates on cognitive trajectories, we conducted a joint test of the linear and quadratic slope to estimate covariate effects on the instantaneous slopes 5, 15, and 25 years after baseline. Women showed greater cognitive decline compared to men at 5, 15, and 25 years after baseline, which was statistically significant 5 years after baseline but was not significant afterwards. The effect of non-Hispanic White was near zero at 5 and 15 years after baseline but was associated with greater decline 25 years after baseline, though this effect did not achieve statistical significance. Additional years of education was associated with a small but significantly faster pace of decline 5 years after baseline, but there was no discernable effect either 15 or 25 years after baseline. Individuals who were older at baseline showed greater, statistically significant decline at all three time points after baseline. When omitting the participants who ever self-reported dementia the overall pattern of results was consistent, although there was no statistically significant effect of sex, education, or age 5 years after baseline. Full results are reported in Table S7.

Discussion

In the HRS-CODA cohort, our results suggest the normative pace of memory decline changes with advancing age, with the pace of change being slower on average at 75 (about −0.04 SD/y), faster at 85 (about −0.08 SD/y), and even faster at 95 (about −0.13 SD/y). An overall pace of about −0.05 SD/y can also be used with caution, as this result comes from a model that shows worse overall fit. These results are within the range of previously identified estimates (−0.01 SD/y to −0.15 SD/y [3, 11–14]). However, this study improves upon limitations of the existing literature through the use of over two decades of follow-up, a narrow age band at baseline, and a nationally representative sample. Moreover, this study improves upon previous research on cognitive aging in the HRS by specifically following the HRS-CODA cohort over time instead of drawing from several different HRS cohorts, which limits the ability to model age heterogeneity and practice/retest effects [31].

Consistent with previous research, this study found an effect of mode of assessment on memory, such that there was better performance on telephone assessments than on those conducted in-person [23, 24]. This may be due to deviations from standard testing protocol when completing a telephone assessment. Age, sex, race/ethnicity, and years of education were all associated with the average level of memory ability (i.e., intercept) in this study. There was some evidence for women and those with greater years of education showing faster decline earlier during follow-up (i.e., at 75). This effect for women was still negative but no longer significant at 85 and 95, whereas the effect for education was both near zero and no longer significant at 85 and 95. Previous research suggests women may have greater cognitive reserve and show faster decline, which could explain the results in this study as well [32]. The reserve effect may also account for higher education being related to greater decline earlier in follow-up, as a recent study of individuals who had experienced a stroke also showed greater education being related to greater cognitive reserve, but steeper decline [33]. Additional research will be needed to explore these findings.

Limitations of this study include the singular cognitive measure used from the HRS. As such, the findings related to memory may not be consistent across other cognitive domains. This study also does not directly model differential survival by demographic characteristics such as sex, race/ethnicity, and socioeconomic status, but the use of full-information maximum likelihood estimation should minimize these biases under the assumption the data are missing at random. The assumption that the data are missing at random cannot be tested using collected data [34], but given the long follow-up period and the inclusion of demographics that are known to be associated with cognition, we do not expect major violations of the missing at random assumption. There is a long prodromal period of dementia as well, which we are not able to capture using these data [35], that may distort the pace of memory decline. Future research should examine changes during the prodromal period, as well as explore cognitive trajectories in other domains of cognition and how this relates to clinical diagnosis, since decline in different domains may be indicative of different forms of dementia [36]. However, domain-specific cognitive measures are not strong in the Core HRS data collection. Research using the more expansive Harmonized Cognitive Assessment Protocol (HCAP) battery in HRS will allow for this kind of work [37]. Moreover, this study is unable to assess period effects, so the trajectory of adults aged 67–74 in 1998 may not generalize to adults aged 67–74 in 2024. The use of the self-reported dementia variable is another limitation, as individuals living with dementia may not be aware of their diagnosis [38].

Despite these concerns, we again note a striking consistency between our findings and those from even older cohorts [3, 39]. The results of this study provide a standardized estimate of normative memory decline that varies by age, but the overall summary is about −0.05 SD/y. Placing our findings in context with the US Study to Protect Brain Health Through Lifestyle Intervention to Reduce Risk (US POINTER) trial, which reported that lifestyle interventions can improve episodic memory by approximately 0.25 SD per year [40], our results suggest that such interventions could offset the expected rate of memory decline by the equivalent of about 6.3 years for individuals in their 70s, about 2.5 years for those in their 80s, and about 1.7 years for those in their 90s. Similar to the results of US POINTER, our findings can serve as a benchmark for future studies on cognitive aging and may be particularly useful for other trials aimed at slowing memory decline to see if the interventions slow memory change to this standard pace among older adults.

Data Availability Statement

Data are available through the Health and Retirement Study. Code used for analyses is available on GitHub: https://github.com/statszach/coda-dr/tree/main.