Midlife hypertension and late‐life cognition: weighting the LifeAfter90 study

Abstract

INTRODUCTION

Midlife hypertension is a well‐established predictor of late‐life cognitive decline; however, few population‐representative estimates of this association exist among oldest‐old cohorts.

METHODS

LifeAfter90 (LA90) is an ethnoracially diverse cohort of individuals aged 90+ linked to midlife multiphasic health check‐ups (MHC). We assessed executive function, semantic memory, and verbal memory at age 90+. We used stabilized inverse probability of censoring weights and linear regression to estimate the association between midlife hypertension and cognitive function in the MHC target population.

RESULTS

After weighting, LA90 participants (n = 307) were similar to the MHC population on race/ethnicity, hypertension, and education. Midlife hypertension was not associated with executive function (β = 0.03, [95% confidence interval [CI]: −0.24, 0.31]) and verbal memory (β = −0.10 [95% CI: −0.37, 0.16]). Midlife hypertension was associated with poorer semantic memory (β = −0.19 [95% CI: −0.44, 0.06]).

DISCUSSION

People with midlife hypertension averaged lower semantic memory using population‐representative models, suggesting that it may have sustained impact on learning and memory in those aged 90+.

Highlights

• Midlife hypertension was associated with late‐life semantic memory.

• Midlife hypertension was not associated with late‐life executive function.

• Midlife hypertension was not associated with late‐life verbal episodic memory.

• Correcting selection biases using weights allowed for population‐level estimates.

1. BACKGROUND

In the 21st century, those ≥90 years old are the fastest growing demographic segment of the US population, ¹ with the burden of age‐related diseases, such as dementia, increasing in parallel. ² , ³ Hypertension is an important risk factor for cognitive change and dementia, due to damage to cerebral blood vessels leading to structural and functional cerebrovascular alterations. ⁴

As the epidemiological evidence accumulates on the effects of hypertension, it reveals a complex relationship that depends, in part, on age at hypertension onset. ⁵ , ⁶ , ⁷ In older adults, a decline in blood pressure (BP) often precedes an Alzheimer's disease diagnosis and predicts a higher risk of dementia; conversely, maintaining higher BP in late‐life appears to be related to preserved cognitive function. ⁶ , ⁸ , ⁹ , ¹⁰ Furthermore, in the 90+ Study, which uses a cohort of individuals 90 years and older, hypertension was associated with lower dementia risk, with a trend toward lower risk as hypertension severity increased. ⁷ These late‐life estimates may be strongly influenced by reverse causation, wherein physiological declines in late‐life BP are a consequence of dementia pathogenesis. ¹¹ Specifically, neurodegeneration may drive reductions in cerebral blood flow via blood–brain‐barrier disruption, leading to microvascular compression ¹² or white matter lesions that induce a state of reduced metabolic demand and tissue atrophy. ¹³ , ¹⁴ In contrast, high hypertension in midlife (40 to 55 years), has been identified as a major risk factor for dementia, cognitive decline, and brain health in late life. ¹⁰ , ¹⁵ , ¹⁶ , ¹⁷ , ¹⁸ Consistent results have come from longitudinal epidemiology studies supporting temporal associations between elevated systolic BP in midlife and late‐life cognitive impairment. ⁶ , ¹⁹ , ²⁰ , ²¹

RESEARCH IN CONTEXT

1. Systematic review: Research shows that midlife hypertension is a risk factor for late‐life cognitive decline, but no estimates of population‐representative dementia incidence exist due to a lack of longitudinal probability samples extending to those 90 years or older. A literature search was conducted using PubMed. The extent to which midlife hypertension impacts cognitive function among those 90+ remains unknown.

2. Interpretation: Those with midlife hypertension had a lower probability of survival until the LA90 enrollment period. We observed that midlife hypertension was associated with poorer semantic memory among the oldest old. Midlife hypertension was not associated with executive function or verbal episodic memory.

3. Future directions: Longitudinal studies of midlife hypertension that examine late‐life neuroimaging biomarkers, including hippocampal volume, are needed to better understand the putative mechanisms and timing between midlife hypertension, neurodegeneration, and memory decline.

However, focusing on midlife measurements introduces the challenge of selective survival. The null results of these studies among the oldest old may be due to self‐selection and survivor bias, well‐known problems in research on older adults. ²² Participants who survive into very old age without dementia and enroll in a study could have systematically different underlying health status than those who die earlier. Individuals who die before age 90 may also have had the most severe hypertension in midlife, because hypertension is an independent risk factor for cardiovascular morbidity and mortality. ²² Therefore, estimates in studies among the oldest old may bias toward the null and the samples may not be representative of target populations of interest – populations for which research is intended to inform policy or clinical decision‐making. ²³ To address this, statistical tools, such as inverse probability of censoring weights (IPCW), aim to correct these biases by accounting for differences in characteristics that affect the outcome of interest between the sample and target population. ²²

Thus far, most studies on the relationship between cognitive decline and midlife hypertension have a short follow‐up of 7 to 20 years from midlife and were based on predominantly White populations. More exploration among the oldest old and diverse populations is critical. ¹⁶ , ¹⁹ , ²⁰ , ²¹ , ²⁴ , ²⁵ , ²⁶ , ²⁷ , ²⁸ Among the few studies incorporating the oldest old (over 85 years), results diverged from the previous midlife literature. ⁹ For example, the Adult Changes in Thought Study found that elevated systolic BP increased dementia risk in the young elderly (65 to 74), but not in those 75 to 84 years or 85 and older. ²⁹ Additionally, the Framingham study among participants ≥80 years old found that midlife hypertension was not associated with late‐life cognitive impairment in memory or executive function. ³⁰

Leveraging data from the LifeAfter90 (LA90) study, we aimed to assess the association between midlife hypertension with late‐life cognitive function among individuals who survive to age 90 years or older. The LA90 study is the largest, most ethnically and racially diverse cohort study of the oldest old. LA90 enrolled members of Kaiser Permanente Northern California (KPNC) members (an integrated healthcare system in the United States), many of whom completed a set of voluntary health surveys and clinical measures as part of multiphasic health check‐ups (MHCs). To estimate population‐representative associations of midlife hypertension on cognitive function among the oldest old, we accounted for the differences in the MHC target population (i.e., individuals born before April 1, 1930, and who had a BP measurement between ages 40 and 55 years and participated in the MHC) and the enrolled LA90 cohort using stabilized IPCW.

2. METHODS

2.1. Samples

We used data from the LA90 cohort and the MHC population. We examined differences in characteristics between the LA90 cohort and the MHC target population to develop IPCWs, allowing us to estimate the mean difference in cognition associated with midlife hypertension among people born before April 1, 1930 (i.e., would have been older than 90 years at time of LA90 enrollment in July 2018) and who reported hypertension status as part of the MHC between the ages of 40 and 55 years.

LA90 is an ongoing cohort study of long‐term members of KPNC, an integrated healthcare delivery system in the United States. Participants are at least 90 years old at the time of enrollment and reside in the San Francisco Bay and Sacramento areas of California. The overall goal of LA90 is to investigate life course contributions to dementia incidence, cognitive decline, neuropathologic changes, and brain imaging markers in an ethnoracially diverse cohort of the oldest old.

KPNC members were eligible for recruitment into the study if they were ≥90 years, KPNC members at any time between 1964 and 1992, and had English or Spanish language proficiency. Participants were excluded if they were unable to provide informed consent, had a medical condition that would interfere with their participation in study interviews (e.g., in hospice care, on dialysis), or if they had a pre‐existing diagnosis of dementia or other neurodegenerative disease in their KPNC electronic medical record at the time of study recruitment. Similar to studies of life course risk factors of dementia among KPNC members, ³¹ , ³² dementia and other neurodegenerative diseases were identified using International Classification of Diseases (ICD) 9 and 10 codes in the electronic health record (described elsewhere). ³³ Participant death details were obtained through KPNC electronic medical records, California State Mortality File, and Social Security Death Records. Study enrollment began in July 2018, and non‐White participants were oversampled to achieve a racially diverse study population. The LA90 study was approved by the KPNC and University of California, Davis, Institutional Review Boards, and all enrolled participants provided informed consent.

We used data from the LA90 cohort and the larger group of MHC participants (who may not have participated in LA90). We consider participants of the MHC representative of the demographic characteristics of the target population living in KPNC regions. ³⁴ The MHCs were a voluntary set of check‐ups offered to KPNC members in San Francisco and Oakland as part of routine care from the 1960s to 1990s. ³⁵ , ³⁶ During the MHCs, a wide range of information was gathered, including information regarding hypertension. These data may include multiple visits per person, and data were harmonized across visits. The MHC has no information about dementia and cannot be used on its own to estimate the mean difference in late‐life cognition associated with midlife hypertension among the target population of MHC participants born before April 1, 1930, who reported hypertension status as part of the MHC between ages 40 and 55 years.

The analytic sample for analyses examining the association between midlife hypertension and late‐life cognition include LA90 participants who participated in at least one MHC visit, including a BP measurement in midlife (40 to 55 years).

2.2. Midlife hypertension assessment

In the primary analyses, midlife hypertension was captured through self‐report during the MHC survey (completed between ages 40 and 55 years) of ever having hypertension, [ever or current] use of antihypertensives or heart medication, or reporting adverse reactions to hypertension medication.

In a sensitivity analysis, we expanded our definition of midlife hypertension to also include clinical measurements of BP obtained from the last MHC visit between the ages of 40 and 55. Following American Heart Association (AHA)/American College of Cardiology (ACC) guidelines, ³⁷ hypertension was defined as a systolic BP of ≥140 mm Hg, a diastolic BP of ≥90 mm Hg, or based on the self‐reported hypertension measurements from the primary analysis.

2.3. Cognitive function assessment

Three cognitive domains (verbal episodic memory, semantic memory, and executive functioning) were assessed by the Spanish and English Neuropsychological Assessment Scales (SENAS) at the LA90 baseline (visit 1). Details of the administration procedures, development, and psychometric characteristics have been extensively described elsewhere. ³⁸ , ³⁹ Briefly, the SENAS was administered during baseline interviews in either English or Spanish, with language of administration determined by an algorithm that considered preferred language and everyday language usage in a variety of settings. The SENAS is a battery of cognitive tests that previously underwent extensive development for valid comparisons of cognitive change across racial/ethnic and linguistically diverse groups. ³⁸ , ³⁹ A verbal episodic memory score was derived from a multi‐trial word‐list‐learning test. Semantic memory was measured using a composite score that was derived from verbal (object‐naming) and non‐verbal (picture association) tests. An executive function composite score was obtained using component tasks of category fluency, phonemic (letter) fluency, and working memory (digit‐span backward, visual‐span backward, list sorting). Each outcome domain was z‐standardized using the baseline sample mean and standard deviation.

2.4. Covariate information

Several covariates were included in the association of midlife hypertension status and cognitive performance in the oldest old. All MHC covariate information was taken from the individual's last MHC visit between 40 and 55 years, except for educational attainment, wherein the highest value ever reported was used. Covariates included MHC centered to the minimum midlife age (40 years), year of MHC centered to the earliest MHC year (1964), sex (male, female), educational attainment (high school diploma or less, trade or technical school, some college, or college degree or higher), and race/ethnicity. To establish the most accurate information at each point in a participant's life, race and ethnicity data were derived either as self‐report from the MHC or from self‐report information in their medical records from the period of LA90 enrollment. For race/ethnicity derived at the time of the MHC, the variable included options for Asian, Black, White, or Other. The “Other” category in MHC‐derived race/ethnicity data primarily encompassed individuals who were later identified in the electronic medical record at the time of LA90 enrollment as Latinx/Hispanic (see Table S1 for full details). For race/ethnicity derived at the time of LA90 enrollment, the options were Asian, Black, Native American, Latinx, or White. We defined race/ethnicity relevant to the study due to the well‐established differences in prevalence and incidence of hypertension ⁴⁰ and dementia ⁴¹ by race and ethnicity. Sex was derived from self‐report or participant medical records. The highest educational attainment ever reported in midlife was extracted from MHC records. Midlife clinical and socioeconomic risk factors were not included as predictors of survival due to the non‐systematic historical capture of the variables during the MHC (15% were missing information on at least one of the following: body mass index, diabetes, or smoking).

2.5. Statistical analysis

We described characteristics for the MHC population and the LA90 study populations using means and standard deviations for continuous variables and frequency (percentages) for categorical variables.

2.5.1. Traditional unweighted analysis

Standard linear regression models with 95% confidence intervals were fitted to estimate domain‐specific baseline cognitive performance in relation to midlife hypertension status. Models were adjusted for the following demographic variables to account for potential confounding due to age‐ and sex‐related physiological changes and period effects: mean‐centered age at MHC, mean‐centered year at MHC, self‐reported race/ethnicity extracted from the medical record, sex, and educational attainment.

2.5.2. Inverse probability weighting analysis

We implemented stabilized inverse probability weights to account for differential participation in the LA90 cohort among the MHC participants. KPNC members who participated in the MHC and would have been age eligible for participation in the LA90 cohort may not have been able to participate in LA90 because they did not survive to the date for enrollment in LA90, the potential enrollee was not a KPNC member at the time of LA90 enrollment despite survival and living in a KPNC region, or they did not enroll in LA90 despite survival and KPNC membership. The possibility that censoring was related to both cognitive function and midlife hypertension exposure (i.e., informative censoring) was addressed using IPCW.

The final inverse selection probability was the product of three weights reflecting the inverse probability of: (1) surviving until LA90 enrollment began (July 1, 2018); (2) being KPNC members at the start of the LA90 enrollment period; and (3) enrolling into LA90 conditional on survival and membership at the start of LA90 enrollment. Weights were calculated using logistic regression models for the probability of remaining uncensored (i.e., propensity score), and adjusted for midlife hypertension, centered age at MHC, centered year at MHC, sex, race/ethnicity, and educational attainment. For weights accounting for differential survival and membership, self‐reported sex, race/ethnicity, and educational attainment were obtained from the MHC. For weights reflecting LA90 enrollment, sex and educational attainment were obtained from the MHC, but race/ethnicity was defined using medical record data. Compared to the MHC race data, the electronic medical records available at the time of LA90 enrollment provided greater granularity (e.g., more specific racial/ethnic categories); therefore, we use this variable for models using individuals who were KPNC members at LA90 enrollment and had electronic medical records available. Each component and the overall final IPCW were stabilized and truncated to the 99th percentile to avoid undue influence of observations with a large weight. ⁴² Linear regression models to estimate the association between midlife hypertension and each cognitive domain were then fit to the pseudopopulation that resulted from the application of the product of the three weights. As in the unweighted models, we adjusted for centered age at MHC, centered year at MHC, sex, medical record race/ethnicity, and educational attainment. For each outcome of executive function, semantic memory, and verbal episodic memory we computed two models: (1) traditional unweighted with self‐reported hypertension and (2) weighted with self‐reported hypertension

2.5.3. Sensitivity analysis

First, we computed weighted models with the expanded hypertension definition of self‐reported hypertension and clinically defined hypertension. Self‐reported hypertension only was utilized as the primary exposure because, during the MHC period (1964 to 1985), self‐report likely captured individuals with more severe or long‐standing disease compared to those identified by a single clinical BP measurement. ⁴³ Second, we repeated our analyses with models for the weights fit with interactions between age at MHC and all the other covariates. Finally, to assess the potential impact of unmeasured midlife clinical and socioeconomic risk factors on the observed associations between midlife hypertension and late life cognitive domains, we conducted quantitative bias analyses by estimating E‐values. ⁴⁴ The E‐value is defined as the minimum strength of association, on the risk ratio scale, that an unmeasured confounder would need to have with both midlife hypertension and late‐life cognition, above and beyond the measured covariates, to explain away an observed association. We report the lower limit of the E‐value confidence intervals to determine the minimum strength of confounding required to shift the observed associations to the null. ⁴⁴ All E‐value calculations were performed using the E‐value package in R. ⁴⁵

All analyses were conducted using R version 4.3.3. In accordance with recommendations for hypothesis‐driven epidemiologic research, ⁴⁶ , ⁴⁷ we did not adjust for multiple comparisons across cognitive domains, as each domain was selected a priori based on distinct etiologic pathways. This approach prioritizes the avoidance of Type II errors in this unique oldest‐old cohort, ensuring that potentially meaningful associations with midlife vascular risk are not masked by alpha‐level penalties. ⁴⁶

3. RESULTS

3.1. LA90 and MHC target population characteristics

There were 307 LA90 participants that were part of the MHC and did not have missing MHC covariates (Table 1). The mean age of the LA90 cohort was 50 (standard deviation [SD] 4.24) years at MHC (Table 1). A total of 109 participants (36%) were male and 50 (32%) were Asian, 99 (32%) were Black, 135 (44%) were White, and 23 (8%) were another race. Overall, 19% of the LA90 cohort self‐reported midlife hypertension (Table 1), and among a larger sample of 433 participants in the sensitivity analysis, 201 (46%) had hypertension as defined by self‐report or through clinical indication. The MHC participants, representing the target population, were similar in average age (50 ± 4.31 years); however, the MHC cohort had a lower proportion of females, a higher proportion of White participants, and higher proportion of participants with no more than a high school education (Table 1).

TABLE 1.

Characteristics of multiphasic health check‐up cohort and LA90 cohort. ^{a .}

	Multiphasic health check‐up (MHC) cohort	LA90 cohort
Characteristics	N = 68,745	N = 307
Survived to LA90 enrollment	23,423 / 68,745 (34%)	—
KPNC members at time of enrollment	5764 / 68,745 (8.4%)	—
Enrolled in LA90	307 / 68,745 (0.4%)	—
Midlife (40 to 55 years) hypertension	—	—
Self‐report	17,708 / 68,745 (26%)	57 / 307 (19%)
Clinical	39,799 / 68,745 (58%)	146 / 307 (48%)
Age at MHC	50.02 (4.31)	50.39 (4.24)
Sex	—	—
Male	32,542 / 68,745 (47%)	109 / 307 (36%)
Female	36,203 / 68,745 (53%)	198 / 307 (64%)
Race b	—	—
Asian	3583 / 68,745 (5.2%)	50 / 307 (16%)
Black	11,489 / 68,745 (17%)	99 / 307 (32%)
Other	3559 / 68,745 (5.2%)	23 / 307 (7.5%)
White	50,114 / 68,745 (73%)	135 / 307 (44%)
Educational attainment	—	—
Grade 0 to 12	29,574 / 67,194 (44%)	90 / 305 (30%)
Technical or business school	6127 / 67,194 (9.1%)	16 / 305 (5.2%)
Partial college	16,736 / 67,194 (25%)	97 / 305 (32%)
College degree or more	14,757 / 67,194 (22%)	102 / 305 (33%)
Missing	1551	2

^a n/N (%); Mean (Standard deviation).

^b Race categories in this table reflect the multiphasic health check‐up data as it was captured at midlife and was available for the entire sample.

3.2. Covariate balance between LA90 and MHC

The covariate balance between the LA90 enrolled cohort and the MHC target population reflects differences in sociodemographic and health characteristics considered to be potential determinants or proxies for determinants of cognitive function both before and after applying the IPCW to the LA90 cohort (Figure 1). Selection patterns are illustrated by covariate balance prior to weighting. For example, prior to weighting (gray circles), LA90 participants were more likely to be White or Black than their MHC counterparts and were more likely to report hypertension. Details regarding the multivariable logistic regression models for the probability of survival until LA90 enrollment, being a KPNC member at the time of LA90 enrollment, and LA90 enrollment are provided in Tables S2–S4. The main predictors of censoring before the LA90 enrollment period (not surviving) were having self‐reported hypertension, older age, Black race/ethnicity, and lower educational attainment (Table S2). The diagnostics of IPCW development, including the propensity score overlap between the LA90 cohort and the MHC sample and distribution of IPCW, appeared reasonable (few extreme values) (Table 2; Figure S1). After developing and applying the IPCW weights to make the LA90 cohort more representative of the MHC target population, the covariate balance was close to a 0.00 standardized mean difference across all variables (Figure 1, black triangles).

FIGURE 1.

Covariate balance comparing the LA90 cohort and MHC target population variables (k) before (circles) and after (triangles) applying stabilized and trimmed inverse probability of censoring weights, characterized by the standardized mean difference. Race/ethnicity categories in this figure (e.g., Latino) reflect the electronic health record data, which provided higher granularity than the original MHC categories shown in Table 1. Standardized mean differences were calculated using the following formula:

[|mean(LA90_k) − mean(MHC _k )|/SD(MHC _k )].

TABLE 2.

Distributions of weights, overall and with stabilization and truncation, for MHC participant survival until LA90 enrollment, for KPNC membership at time of LA90 enrollment, LA90 enrollment, and final weight.

Weighting scheme	Survival until enrollment	KPNC membership	LA90 enrollment	Final weight a
Weights (no adjustment)	—	—	—	—
Minimum	1.097	1.011	1.006	—
25%	1.348	1.085	1.026	—
Median	1.651	1.253	1.044	—
75%	2.201	1.924	1.079	—
Maximum	11.245	46.362	58.515	—
Stabilized and truncated weights	—	—	—	—
Minimum	0.459	0.282	0.139	0.131
25%	0.813	0.782	0.970	0.301
Median	0.923	0.839	0.985	0.445
75%	1.115	1.042	1.015	0.778
Maximum	2.207	3.022	1.409	4.495

Abbreviations: KPNC, Kaiser Permanente Northern California; LA90, LifeAfter90; MHC, multiphasic health check‐up.

^a Product of the three weight schemes and applied to the association between midlife hypertension and the three cognitive performance measures of executive function, semantic memory, and verbal episodic memory.

3.3. Estimates of associations between self‐reported midlife hypertension and cognitive function domains from conventional model

There were no associations between self‐reported midlife hypertension and domain‐specific baseline cognitive function in conventional models, adjusting for centered age at MHC, centered year at MHC, sex, medical record race/ethnicity, and educational attainment (Figure 2).

FIGURE 2.

Unweighted and weighted linear regressions testing the association between midlife hypertension and cognitive scores for executive function, semantic memory, and verbal episodic memory from the Spanish and English Neuropsychological Assessment Scales (SENAS). Circles represent the unweighted results, triangles represent the main analysis results using self‐reported hypertension, antihypertensive medication, or adverse reaction to hypertensive medication, and diamonds represent the addition of a clinical definition to the overall hypertension definition.

3.4. Estimates of associations between self‐reported midlife hypertension and cognitive function domains from weighted model

There were no associations between self‐reported midlife hypertension and baseline executive function or verbal episodic memory in weighted models (Figure 2). Self‐reported midlife hypertension was associated with lower baseline semantic memory scores (β = −0.19; 95% confidence interval [CI] [−0.44, 0.06]).

3.5. Sensitivity analyses

There were no associations between the expanded midlife hypertension definition that includes clinical measures of elevated BP and baseline executive function or verbal episodic memory in weighted models (Figure 2). On the other hand, the expanded midlife hypertension definition was associated with lower baseline semantic memory scores (β = −0.22; 95% CI [−0.39, −0.04]) compared to those with normotensive BP. Our results from the second sensitivity analysis that included weights with interactions between age at MHC and all the other covariates were similar to the main analyses (Table S5). The distribution of weights from interaction models appeared reasonable (Figure S2). Finally, to assess the potential impact of unmeasured confounders on our findings, we calculated E‐values for the association between midlife hypertension and the cognitive domains (Table S6). In our primary weighted model for midlife hypertension and semantic memory performance, we found an E‐value of 1.68 (95% CI lower bound: 1.0) and a value of 1.76 (95% CI lower bound: 1.0) for the weighted model, including clinical and self‐reported measures of BP.

4. DISCUSSION

In this prospective cohort study of racially and ethnically diverse oldest‐old adults, we found that midlife hypertension was associated with lower semantic memory scores in late life when compared with participants with midlife normotension and observed no differences in executive function or verbal episodic memory scores. The LA90 cohort and the MHC participants varied by race and education and the impact of the unweighted and weighted results shifted across cognitive domain outcomes, with a dramatic impact on semantic memory from a null association to significantly lower scores in hypertension compared to normotension. Across each outcome, including a definition of clinical hypertension along with the self‐reported midlife measure allowed for similar, though more precise, estimates due to the larger sample size.

Overall, our findings suggest that midlife hypertension may be an important contributor to semantic memory performance among those aged 90+ years, but not to executive function or verbal episodic memory. This study is one of the few to examine how midlife hypertension may be associated with domain‐specific late‐life cognitive function among the oldest old. Our findings are consistent with prior literature suggesting that midlife hypertension is associated with cognitive decline in memory in younger older adults. ⁴⁸ , ⁴⁹ , ⁵⁰ In particular, our estimates for the association between midlife hypertension and semantic memory align with the direction of effect found in a recent a systematic review of 10 longitudinal cohorts, where eight found negative relationships with memory (mean difference: −0.06, 95% CI [−0.20, 0.08]). ⁵⁰ While Joyce et al. ⁵⁰ reported associations with memory as a broad domain, our study suggests this relationship may be driven specifically by semantic memory in the oldest old. In addition, the lack of observed associations with executive function and episodic memory in our sample contrasts with the findings from the younger‐old cohorts in the review. ⁵⁰ This contrast may reflect the unique survivor profile of our 90+ cohort.

Overall, our associations shifted slightly closer to the null when the definition of hypertension included self‐report of hypertension and a clinical definition of hypertension. We believe that during midlife, when the LA90 participants participated in the MHC exams between 1964 and 1985, the subset of individuals self‐reporting hypertension likely had more severe disease than those with an indication of high BP found using the AHA/ACC guidelines. ³⁷ This is because public and professional awareness regarding detection, evaluation, treatment, and control of hypertension was steadily increasing over the same 20‐year period. ⁴³ These results highlight the importance of the exact definition of midlife hypertension in the context of secular trends and its associations with domain‐specific cognitive function in oldest‐old populations.

Past research ⁵⁰ has found that specific domains, such as memory, are more affected than others by midlife hypertension. The relationship is especially pronounced among those whose hypertension that began in midlife is sustained for decades (25 to 30 years). The hippocampus and entorhinal cortex – structures essential for learning and memory – are particularly vulnerable to anatomical and pathophysiological changes associated with cardiovascular risk factors, including vascular remodeling, white matter lesions, and amyloid angiopathy. ⁵¹ , ⁵² , ⁵³ Research indicates that some semantic memory processing declines with age, potentially due to reduced processing efficiency or central‐executive functioning that support efficient semantic retrieval ⁵⁴ in the presence of cardiovascular changes. ⁵⁵ However, current evidence linking brain imaging markers affected by hypertension to semantic memory declines, remains mixed. Some longitudinal studies report modest associations with white matter hyperintensities, while others report null findings. ⁵⁶ , ⁵⁷ , ⁵⁸ Overall, while we observed lower semantic memory function in late life, executive function was unaffected, suggesting that midlife hypertension may have selective long‐term impacts on the mechanisms linking to cognitive function that should be further explored.

This study has several limitations. Only a subset of LA90 participants had MHC self‐reported hypertension or clinical measurements, limiting the sample size of the LA90 cohort to 307 individuals. Furthermore, when using IPCW, we were limited in the number of covariates we could include as incorporating more variables from the MHC would have increased missingness and reduced the effective sample size. Specifically, while midlife comorbidities such as diabetes and body mass index are relevant to late‐life cognition, these data were not systematically captured for all MHC participants during the period 1964 to 1985; their inclusion would have introduced selection bias by excluding individuals with missing information. The validity of our weighted estimates relies on the assumption that the factors predicting survival, being a KPNC member at the start of enrollment, and enrollment in the LA90 cohort are well captured by our available covariates. We acknowledge that due to missingness of some midlife comorbidities, our weight models may not fully account for all selection pressures. This could result in residual selection bias if the unmeasured factors are strong confounders in the models. However, our use of E‐values provides a quantitative assessment of how robust our observed associations are to such unmeasured factors. Although the lower confidence limits for our E‐values were 1.0, likely reflecting the small sample size of the LA90 analytic population, our E‐value point estimates suggest that the observed associations may be robust to potential unmeasured confounding from these factors. The sample size also restricted our ability to investigate racial/ethnic or sex differences in the association between midlife hypertension and domain‐specific cognitive performance among the oldest old. An examination of effect measure modification by demographic differences with small sample sizes would have resulted in greater variances in the estimated stabilized weights, leading to large uncertainties around the weighted cognitive performance estimates.

Our work is the first study of cognitive domain‐specific function among a diverse cohort of individuals 90+ that evaluates and accounts for differences in characteristics of the sample and a target population (e.g., long‐term members of the KPNC system represented by individuals who attended MHCs). Future work with the LA90 cohort and MHC population may be used to estimate population‐representative associations from midlife hypertension for dementia incidence rates and cumulative risk. Our investigation into the burden of midlife hypertension on cognitive function among a diverse oldest‐old cohort informs etiologic research by revealing whether vascular risks found in midlife indeed have a longitudinal reach into the tenth decade that are masked by survival bias in late‐life cohorts. This investigation is increasingly relevant as US life expectancy grows and the population ages. ¹ Those 90 years and above are projected to grow to from approximately 2.4 million in 2024 to over 17 million by 2100, making it the fastest growing segment of the US population. ¹ , ⁵⁹ Current population epidemiology suggests that the ≥90 age group has the highest rates of dementia and disability, ³ , ⁶⁰ posing an urgent public health challenge that can be addressed using life course data. Our results are bolstered by weight diagnostics that show no evidence of extreme weights or positivity violations that would indicate instability. This suggests that our weighted estimates are the result of statistical corrections for survival and enrollment patterns of the MHC target population that were otherwise masked in the unweighted sample. Given the current lack of any population‐representative samples with detailed data for cognitive function among multiethnic individuals aged 90+, generalizability and transportability methods are needed to both evaluate the external validity of studies and inform public health and policy work. Finally, our study highlights key methodological challenges when using weighting to extend inferences from observational samples to target populations, including variable selection, model specification, and evaluating conditional exchangeability, particularly in studies with small sample sizes.

In conclusion, we assessed differences between an ethnoracially diverse cohort of long‐term KPNC members and a population of individuals who attended MHCs and used stabilized IPCW to examine the association between midlife hypertension and cognitive function among the oldest old. We found that unweighted and weighted results were similar across each cognitive domain, except for semantic memory, in which the weighted estimates show that midlife hypertension is associated with lower semantic memory performance. Our results emphasize the importance of measuring variables in both specialized cohorts and population‐level datasets to support the assessment of external validity and transportability, especially in studies of aging.

CONFLICT OF INTEREST STATEMENT

The authors declare no conflicts of interest. Author disclosures are available in the Supporting Information.

CONSENT STATEMENT

The study was approved by the Kaiser Permanente Northern California and University of California at Davis Institutional Review Boards, and all enrolled participants provided informed consent.

Supporting information

Supporting Information

Supporting Information