Education, Incident Cancer, And Rate of Memory Decline in a National Sample of US Adults in Mid-to-Later-Life

Abstract

Introduction:

Middle-aged and older adults who develop cancer experience memory loss following diagnosis, but memory decline in the years before and after cancer diagnosis is slower compared to their cancer-free counterparts. Educational attainment strongly predicts memory function during aging, but it is unclear whether education protects against memory loss related to cancer incidence or modifies long-term memory trajectories in middle-aged and older cancer survivors.

Materials and Methods:

Data were from 14,449 adults (3,248 with incident cancer, excluding non-melanoma skin cancer) aged 50+ in the population-based US Health and Retirement Study from 1998 to 2016. Memory was assessed every two years as a composite of immediate and delayed word recall tests and proxy assessments for impaired individuals. Memory scores all time points were standardized at to the baseline distribution. Using multivariate-adjusted linear mixed-effects models, we estimated rates of memory decline in the years before cancer diagnosis, shortly after diagnosis, and in the years after diagnosis. We compared rates of memory decline between incident cancer cases and age-matched cancer-free adults, overall and according to level of education (<12 years, “low”; 12 to <16 years, “intermediate”; ≥16 years, “high”).

Results:

Incident cancer diagnoses were followed by short-term declines in memory averaging 0.06 standard deviation (SD) units (95% confidence interval [CI]: −0.084, −0.036). Those with low education experienced the strongest magnitude of short-term decline in memory after diagnosis (−0.10 SD units, 95% CI: −0.15, −0.05), but this estimate was not statistically significantly different from the short-term decline in memory experienced by those with high education (−0.04 SD units, 95% CI: −0.08, 0.01; p-value for education as an effect modifier = 0.15). In the years prior to and following an incident cancer diagnosis, higher educational attainment was associated with better memory, but it did not modify the difference in rate of long-term memory decline between cancer survivors and those who remained cancer-free.

Conclusions:

Education was associated with better memory function over time among both cancer survivors and cancer-free adults aged 50 and over. Low education may be associated with a stronger short-term decline in memory after a cancer diagnosis.

1. Introduction

Cognitive decline is a negative outcome often experienced by older cancer survivors, which has been attributed to cancer treatments and psychological stress. ^1–6 Emerging results have further demonstrated that some chemotherapies may have neurotoxic effects.^4,7–11 However, there is limited evidence on factors that modify the risk of cancer-related cognitive decline. Education, a commonly used proxy marker of cognitive reserve, is thought to allow individuals to maintain cognitive function in the face of neurological insults.^12–16 Education has been studied extensively as a risk-modifying factor for Alzheimer’s disease and related dementias (ADRD).^17–26 In the context of ADRD, education is thought to allow individuals to maintain cognitive function despite the presence of brain pathology, through mechanisms such as recruiting alternative neural networks or employing pre-existing cognitive processing approaches.^14,27

To the best of our knowledge, no prior work has brought together these two bodies of evidence to evaluate whether the benefits of education for ADRD are also relevant for cancer-related cognitive decline. While ADRD is understood to have different causes than cancer-related cognitive decline, evidence on whether education modifies risk for cancer-related cognitive decline, as it does for ADRD, will indicate whether the mechanisms of cognitive reserve may apply to non-ADRD cognitive problems and provide information of direct relevance to patients with cancer. If education is protective against cancer-related cognitive decline in the short- or long-term, it may help predict cognitive outcomes for cancer survivors and point the way towards tailored interventions to promote cognitive well-being.^28–30

Most existing studies on cancer-related cognitive decline have controlled for the effects of education using statistical modeling (i.e., as a confounder),^31–36 with few directly investigating the rates of cancer-related cognitive decline across levels of education (i.e., as an effect modifier).^35,37 These studies suggest that older cancer survivors with higher education may experience less severe cancer-related cognitive decline than those with lower education.^35,37 However, this existing research has been conducted in clinic-based settings. While such settings allow for detailed clinical data to be collected, clinic-based studies of cancer survivorship are often non-representative of the general population of cancer survivors and may over-represent highly educated participants.^38–40 They thus often have insufficient variation in educational attainment to investigate it as a risk-modifying factor for cancer-related cognitive decline.

Population-representative samples may more accurately capture the distribution of educational attainment in the general population.⁴¹ Further, population-based cohorts provide longitudinal data prior to cancer diagnosis, which is typically unavailable in clinic-based studies that recruit participants shortly after a cancer diagnosis.^{2,31–34,42–45} Pre-diagnosis data on cognitive outcomes allows the observation of within-person cognitive aging trajectories and the comparison of these trajectories between cancer survivors to those of individuals who remain cancer-free.^46–48 Factors that modify risk for cancer-related cognitive decline might become evident from analyses of population-representative cohorts with longitudinal pre- and post-diagnostic data on cognitive outcomes and cancer-free comparison groups.^40,46

To address these gaps in the literature and contribute evidence using a large, population-based sample with longitudinal, within-person data both pre- and post-cancer diagnosis, we used data from the US Health and Retirement Study (HRS) over an 18-year period from 1998 to 2016 to investigate the role of educational attainment as a risk-modifying factor for cancer-related memory decline. We hypothesized that higher educational attainment would ‘buffer’ any short-term negative changes in memory function experienced after a cancer diagnosis.

2. Materials and Methods

2.1. Study Design and Sample Population

Data were from biennial interviews in the population-based US HRS.⁴¹ Participants eligible for the analysis were born before 1949, completed interviews in 1998 (baseline), had no history of cancer at baseline, and at least one follow-up assessment after baseline. Full details of the study design have been described elsewhere.⁴⁹ Data on incident cancer, memory function, and covariates were assessed biennially from 1998 through 2016 via telephone and in-person interviews. For HRS participants who were too impaired to directly participate, study interviews were conducted with proxy respondents, typically spouses or other family members.⁵⁰ The final analytical sample was 14,449 participants (Figure 1).

Figure 1. Flow diagram for sample selection.

The HRS was approved by the University of Michigan Health Sciences and Behavioral Sciences Institutional Review Board and these analyses were determined exempt from review by the University of California, San Francisco Institutional Review Board.

2.2. Measures

2.2.1. Incident Cancer

Incident cancer was assessed as a new self-reported physician diagnosis of a cancer or malignant tumor, excluding non-melanoma skin cancers (N=3,248). Month and year of diagnosis were collected. To retain participants with an incomplete date of diagnosis (n = 97), we used the midpoint between the last cancer-free interview date and the date when the first cancer was reported. For participants who died during follow-up (N=6,672), we used post-death proxy interviews (a.k.a. exit interviews) to ascertain incident cancer diagnosis status prior to death (N=861).

2.2.2. Memory

Memory was assessed as immediate and delayed recall of a 10-word list administered during each biennial study interview. To retain severely impaired participants and reduce possible bias introduced by their exclusion, we incorporated proxy assessments of the participants’ memory. We used a validated memory algorithm which combines direct and proxy assessments of memory into a composite memory score.⁵⁰ This algorithm was not validated for Hispanic respondents, so our analyses are restricted to non-Hispanic respondents (of any self-identified race). We standardized memory scores at each follow-up time point using the baseline sample mean and standard deviation (SD), therefore analyses present rates of memory change relative to the baseline distribution.

2.2.3. Education

Education was assessed by self-reported years of schooling at baseline. For analyses, we coded years of education continuously, mean-centered at 12 years, and with all categories higher than a bachelor’s degree classified as 17 years of education. We also conducted analyses using the education variable categorized as “low”: <12 years, “intermediate”: 12 to <16 years, and “high”: ≥16 years.

2.2.4. Covariates

Potential confounders of the relationship between incident cancer and rate of memory change were: self-reported sex (male vs. female), race (White vs. non-White), southern US birthplace (Southern vs. non-South), childhood socioeconomic status (cSES) index score,⁵¹ baseline household wealth (natural log scale), self-rated childhood health using a using a 5-item Likert scale (categorized as poor/fair, good, and, very good/excellent health), baseline measures of vigorous physical activity, baseline smoking status (current, ever, or never smoker), baseline alcohol use (none, low risk, and binge definitions from the National Institute on Alcohol Abuse and Alcoholism⁵²), height (continuous, in meters, as a measure of childhood net nutrition^53–55), baseline body mass index (centered at 25), and self-reported baseline comorbidities (hypertension, diabetes, stroke, heart disease, and lung disease). Cancer stage/grade and cancer treatments were not included in the analysis as these variables lie on the causal pathway between cancer and memory change.⁵⁶

2.3. Statistical Analysis

Differences in baseline characteristics between participants with and without a new cancer over the follow-up by educational levels were described with Pearson χ² statistics for categorical variables and t tests for normally distributed continuous variables. We used multivariable linear mixed-effects models to estimate longitudinal memory trajectories by incident cancer status, overall and according to education. The primary timescale, age at each study interview, described the average rate of memory decline per decade of age for participants with no cancer diagnosis during the entire follow-up period. To estimate differences in memory over time for participants who developed cancer, we specified four cancer-related terms in the model. The first was a binary cancer diagnosis indicator variable for whether a participant had been diagnosed with cancer over the follow-up. The second was a time-dependent cancer indicator variable, set to 0 for all time points prior to cancer diagnosis, and 1 for all time points at and after diagnosis. This time-dependent cancer indicator variable was used to estimate the short-term change in memory following diagnosis. The third and fourth cancer-related model terms were two cancer slope terms that described differences in long-term memory trajectories in the periods before and after diagnosis, compared to cancer-free individuals. The pre-cancer period was assigned negative time (in decades) until the date of diagnosis, and 0 for all time points after diagnosis. The post-cancer period was assigned positive time (in decades) after diagnosis, and 0 for all time points before diagnosis. Participants who did not report cancer over the follow-up were assigned values of zero for pre- and post-cancer period. We centered age at each interview at 75 for individuals without cancer, and age at diagnosis at 75 for individuals with incident cancer. Centering the age variables at 75 allowed us to align the memory decline slopes for individuals with and without cancer, and to estimate the effect of a cancer diagnosis on memory decline over and above aging alone. A quadratic term for age at each interview was included to estimate curvilinear trajectories of memory. We adjusted models for baseline covariates and interactions between covariates and age at each interview.

To estimate memory slopes before and after cancer diagnosis according to education, we first estimated a model with interaction terms between continuous years of education and each of the four cancer-related variables. This model allowed us to test for effect modification of the cancer-memory relationships (both the long-term pre- and post-diagnosis memory decline slopes, and the short-term change in memory following diagnosis) by each additional year of education.

Models with continuous years of education should maximize statistical power if the association is approximately dose-response, but coefficients are difficult to interpret and will misrepresent patterns if there is a marked threshold in the effects.^57,58 To address this issue, we next investigated whether effect modification by educational attainment was present when individuals achieved certain educational levels associated with credentials such as a high school diploma or college degree. We thus specified models stratified by each of the three educational attainment categories. To test for effect modification by these educational attainment categories, we specified a single model using the full sample with interaction terms between each educational category and each of the four cancer-related variables. All models allowed participants’ within-person age slopes and intercepts to vary as random effects.⁵⁹ The covariance structure of random effects was coded as unstructured, allowing the correlation between intercepts and slopes to be estimated, and restricted maximum likelihood estimation (REML) was used. All analyses were conducted using Stata/SE version 15.1 (StataCorp, College Station, TX).

2.4. Sensitivity Analysis:

We identified participants with no memory assessments after their reported incident cancer diagnosis (n=234 in the low education category, n=374 in the intermediate education category, and n=101 in the high education category). To determine the potential impact of these individuals on our estimates, we performed a sensitivity analysis restricting the cancer survivors in our sample to those with at least one post-diagnosis memory assessment.

3. Results

3.1. Characteristics of the Sample

Regardless of educational levels, participants with a cancer diagnosis were more often male, taller, and reported a higher frequency of tobacco and alcohol use. The prevalence of baseline comorbidities did not differ between those with and without a cancer diagnosis over the follow up (Supplemental Table 1). Baseline characteristics of study participants are shown in Table 1. Mean age at baseline was 65.9 years (SD 10.0 years). Participants with low education were older than those with intermediate and high education, regardless of cancer status. The median number of follow-up assessments for those with low education was 6 (range 2–10), for those with intermediate education was 8 (range 2–10), and for those with high education was 9 (range 2–10). The number of study assessments after a cancer diagnosis ranged from 0–10 (median 2). Proxy assessments of memory function were more frequent in participants with lower educational levels (Supplemental Table 2). The frequency of proxy memory assessments increased over the follow up, and this happened more frequently in participants with a cancer diagnosis than those without cancer.

Table 1.

Baseline characteristics of study participants by education and cancer diagnosis status, the Health and Retirement Study, United States, 1998–2016

Baseline Characteristic	Total n = 14 449	Educational attainment
		Less than 12 years		12 to <16 years		16 or more years
		Cancer-free	Incident cancer	Cancer-free	Incident cancer	Cancer-free	Incident cancer
		n = 2 926	n = 843	n = 6 171	n = 1 747	n = 2 104	n = 658
Age, mean (SD), y	65.9 (10.0)	69.4 (10.6)	67.9 (8.9)	65.0 (9.9)	65.3 (8.8)	63.6 (9.7)	64.0 (8.3)
Male, n (%)	6,057 (41.9)	1,126 (38.5)	460 (54.6)	2,172 (35.2)	804 (46.0)	1,098 (52.2)	397 (60.3)
White, n (%)	11,994 (83.0)	2,049 (70.0)	603 (71.5)	5,345 (86.6)	1,533 (87.8)	1,856 (88.2)	608 (92.4)
Childhood SES index, mean (SD)	0.075 (0.86)	−0.39 (0.74)	−0.39 (0.76)	0.130 (0.78)	0.105 (0.78)	0.57 (0.90)	0.58 (0.89)
Southern birthplace, n (%)	5,308 (36.7)	1,565 (53.5)	462 (54.8)	1,982 (32.1)	518 (29.7)	606 (28.8)	175 (26.6)
Household wealth, per $10,000, median (IQR)	14.1 (4.8 34.3)	5.4 (0.713.9)	6.1 (1.015.2)	15.1 (6.034.5)	16.0 (6.432.8)	32.1 (14.071.3)	34.8 (15.777.6)
Height, mean (SD), m	1.69 (0.1)	1.67 (0.1)	1.70 (0.1)	1.68 (0.1)	1.70 (0.1)	1.71 (0.1)	1.74 (0.1)
BMI, mean (SD), kg/m2	27.0 (5.2)	27.4 (5.6)	27.5 (5.2)	27.0 (5.2)	27.1 (5.0)	26.3 (4.7)	26.6 (4.8)
Vigorous physical activity, n (%)	6,535 (45.2)	1037 (35.4)	341 (40.5)	2,912 (47.2)	844 (48.3)	1,055 (50.1)	346 (52.6)
Current smoking	2,402 (16.6)	507 (17.3)	228 (27.0)	1,002 (16.3)	372 (21.3)	211 (10.0)	78 (11.9)
Ever smoking	8,500 (58.8)	1,709 (58.4)	606 (71.9)	3,479 (56.4)	1,158 (66.3)	1,145 (54.4)	403 (61.2)
Alcohol use, n (%)
None	10,008 (69.3)	2451 (83.8)	635 (75.3)	4,326 (70.1)	1,138 (65.1)	1,124 (53.4)	334 (50.8)
Low	4,169 (28.9)	422 (14.4)	175 (20.8)	1,734 (28.1)	563 (32.2)	958 (45.5)	317 (48.2)
Binge	272 (1.9)	53 (1.8)	33 (3.9)	111 (1.8)	46 (2.6)	22 (1.1)	7 (1.1)
Self-rated childhood health, n (%)
Excellent/V ery good	10,928 (75.6)	1902 (65.0)	549 (65.1)	4,798 (77.8)	1,377 (78.8)	1,740 (82.7)	562 (85.4)
Good	2,645 (18.3)	754 (25.8)	217 (25.7)	1,058 (17.1)	280 (16.0)	267 (12.7)	69 (10.5)
Poor/Fair	876 (6.1)	270 (9.2)	77 (9.1)	315 (5.1)	90 (5.2)	97 (4.6)	27 (4.1)
Hypertensio n, n (%)	6,064 (42.0)	1501 (51.3)	403 (47.8)	2,498 (40.5)	719 (41.2)	715 (34.0)	228 (34.7)
Diabetes, n (%)	1,681 (11.6)	504 (17.2)	122 (14.5)	644 (10.4)	183 (10.5)	179 (8.5)	49 (7.4)
Heart disease, n (%)	2,755 (19.1)	768 (26.2)	212 (25.1)	1,082 (17.5)	325 (18.6)	278 (13.2)	90 (13.7)
Stroke, n (%)	890 (6.2)	301 (10.3)	69 (8.2)	323 (5.2)	102 (5.8)	70 (3.3)	25 (3.8)
Lung disease, n (%)	904 (6.3)	269 (9.2)	95 (11.2)	333 (5.4)	126 (7.2)	56 (2.7)	25 (3.8)

Note: The cancer-free and incident cancer groups refer to cancer status by the end of the follow-up period in 2016

Abbreviations: BMI, body mass index; IQR, interquartile range; m, meters; SD, Standard deviation; SES, socioeconomic status.

3.2. Effect modification of short-term and long-term post-diagnosis decline in memory by continuous years of education:

Participants without a cancer diagnosis during the follow-up and a high school education (i.e., 12 years of education) experienced an average aging-related memory decline of 0.81 SD units (95% CI: −0.84, −0.85) in the decade between age 65 and 75 years, and 1.47 SD units (95% CI: −1.50, 1.37) in the decade between age 75 and 85 years. Each additional year of education was associated with a rate of memory decline that was 0.018 SD units slower per decade (95% CI: 0.015, 0.022). Estimated coefficients and the equation used to calculate these outcomes are in Table 2.

Table 2.

Estimated Effects of Incident Cancer and Education on Memory Function and Annual Rate of Memory Change using Linear Mixed-Effects Models, the US Health and Retirement Study, United States, 1998–2016.

Characteristic	β	Model 1a 95% CI	P-value
Memory function (SD units) and memory change (SD units/decade) in those with 12 years of education
Participants with no cancer during follow-up:
Memory function at age 75	−0.894	(−0.937, −0.851)	<0.001
Memory slope with linear age (centered at 75)	−1.111	(−1.140, −1.077)	<0.001
Memory slope with quadratic age (SD units /decade2)	−0.298	(−0.303, −0.293)	<0.001
Participants with an incident cancer diagnosis:
Difference in memory score right before cancer diagnosisb	0.052	(0.024, 0.080)	<0.001
Short-term change in memory following diagnosis	−0.060	(−0.084, −0.036)	<0.001
Difference in memory slope before diagnosisb	0.047	(0.021, 0.074)	<0.001
Difference in memory slope after diagnosisb	0.058	(0.019, 0.098)	0.004
Effect of one additional year of education on the following estimates:
Participants with no cancer during follow-up:
Memory function at age 75	0.050	(0.045, 0.054)	<0 .001
Memory slope with linear age	0.018	(0.015, 0.022)	<0.001
Participants with an incident cancer diagnosis:
Difference in memory score right before cancer	−0.003	(−0.010, 0.004)	0 .361
Short-term change in memory following diagnosis	0.005	(−0.030, 0.013)	0.219
Difference in memory slope before diagnosis	−0.006	(−0.014, 0.002)	0.149
Difference in memory slope after diagnosis	0.002	(−0.010, 0.014)	0.711

^aModel adjusted for sex, race, southern birthplace, childhood socioeconomic index, total household wealth, height, body mass index, vigorous physical activity, current and ever smoking, alcohol use, childhood self-rated health, and comorbidities (hypertension, diabetes, heart disease, stroke, lung disease)

^bCompared to participants with no cancer over the follow-up, as the reference group $$\begin{gathered} \mathrm{M}\mathrm{e}\mathrm{m}\mathrm{o}\mathrm{r}{\mathrm{y}}_{\mathrm{i}\mathrm{j}}=\mathrm{\beta }0+\mathrm{\beta }1\mathrm{c}\mathrm{a}\mathrm{n}\mathrm{c}\mathrm{e}\mathrm{r}\_\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{g}\mathrm{n}\mathrm{o}\mathrm{s}\mathrm{i}\mathrm{s}\_\mathrm{d}\mathrm{u}\mathrm{r}\mathrm{i}\mathrm{n}\mathrm{g}\_\mathrm{f}\mathrm{o}\mathrm{l}\mathrm{l}\mathrm{o}\mathrm{w}\_\mathrm{u}{\mathrm{p}}_{\mathrm{i}}+\mathrm{\beta }2\mathrm{c}\mathrm{a}\mathrm{n}\mathrm{c}\mathrm{e}\mathrm{r}\_\mathrm{n}\mathrm{o}{\mathrm{w}}_{\mathrm{i}\mathrm{j}}+\mathrm{\beta }3\mathrm{t}\mathrm{i}\mathrm{m}\mathrm{e}\_\mathrm{t}\mathrm{o}\_\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{g}\mathrm{n}\mathrm{o}\mathrm{s}\mathrm{i}{\mathrm{s}}_{\mathrm{i}\mathrm{j}}(\mathrm{z}\mathrm{e}\mathrm{r}\mathrm{o}\mathrm{f}\mathrm{o}\mathrm{r}\mathrm{t}\mathrm{i}\mathrm{m}\mathrm{e} \\ \mathrm{a}\mathrm{f}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{c}\mathrm{a}\mathrm{n}\mathrm{c}\mathrm{e}\mathrm{r})+\mathrm{\beta }4\mathrm{t}\mathrm{i}\mathrm{m}\mathrm{e}\_\mathrm{s}\mathrm{i}\mathrm{n}\mathrm{c}\mathrm{e}\_\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{g}\mathrm{n}\mathrm{o}\mathrm{s}\mathrm{i}\mathrm{s}{\mathrm{i}}_{\mathrm{i}\mathrm{j}}(\mathrm{z}\mathrm{e}\mathrm{r}\mathrm{o}\mathrm{f}\mathrm{o}\mathrm{r}\mathrm{t}\mathrm{i}\mathrm{m}\mathrm{e}\mathrm{b}\mathrm{e}\mathrm{f}\mathrm{o}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{n}\mathrm{c}\mathrm{e}\mathrm{r})+\mathrm{\beta }5\mathrm{a}\mathrm{g}{\mathrm{e}}_{\mathrm{i}\mathrm{j}}+\mathrm{\beta }6\mathrm{a}\mathrm{g}{\mathrm{e}}_{\mathrm{i}\mathrm{j}}^2+\mathrm{\beta }7 \\ \mathrm{a}\mathrm{g}\mathrm{e}\_\mathrm{a}\mathrm{t}\_\mathrm{c}\mathrm{a}\mathrm{n}\mathrm{c}\mathrm{e}\mathrm{r}\_\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{g}\mathrm{n}\mathrm{o}\mathrm{s}\mathrm{i}{\mathrm{s}}_{\mathrm{i}}+\mathrm{\beta }8\mathrm{e}\mathrm{d}\mathrm{u}\mathrm{c}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}{\mathrm{n}}_{\mathrm{i}}+\mathrm{\beta }9\mathrm{e}\mathrm{d}\mathrm{u}\mathrm{c}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}{\mathrm{n}}_{\mathrm{i}}{}^{*}\mathrm{a}\mathrm{g}{\mathrm{e}}_{\mathrm{i}\mathrm{j}}+\mathrm{\beta }10 \\ \mathrm{e}\mathrm{d}\mathrm{u}\mathrm{c}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}{\mathrm{n}}_{\mathrm{i}}*\mathrm{c}\mathrm{a}\mathrm{n}\mathrm{c}\mathrm{e}\mathrm{r}\_\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{g}\mathrm{n}\mathrm{o}\mathrm{s}\mathrm{i}\mathrm{s}\_\mathrm{d}\mathrm{u}\mathrm{r}\mathrm{i}\mathrm{n}\mathrm{g}\_\mathrm{f}\mathrm{o}\mathrm{l}\mathrm{l}\mathrm{o}\mathrm{w}\_\mathrm{u}{\mathrm{p}}_{\mathrm{i}}+\mathrm{\beta }11\mathrm{e}\mathrm{d}\mathrm{u}\mathrm{c}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}{\mathrm{n}}_{\mathrm{i}}*\mathrm{c}\mathrm{a}\mathrm{n}\mathrm{c}\mathrm{e}\mathrm{r}\_\mathrm{n}\mathrm{o}{\mathrm{w}}_{\mathrm{i}\mathrm{j}}+\mathrm{\beta }12 \\ \mathrm{e}\mathrm{d}\mathrm{u}\mathrm{c}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}{\mathrm{n}}_{\mathrm{i}}^{*}\mathrm{t}\mathrm{i}\mathrm{m}\mathrm{e}\_\mathrm{t}\mathrm{o}\_\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{g}\mathrm{n}\mathrm{o}\mathrm{s}\mathrm{i}{\mathrm{s}}_{\mathrm{i}\mathrm{j}}+\mathrm{\beta }13\mathrm{e}\mathrm{d}\mathrm{u}\mathrm{c}\mathrm{a}\mathrm{t}\mathrm{i}\mathrm{o}{\mathrm{n}}_{\mathrm{i}}^{*}\mathrm{t}\mathrm{i}\mathrm{m}\mathrm{e}\_\mathrm{s}\mathrm{i}\mathrm{n}\mathrm{c}\mathrm{e}\_\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{g}\mathrm{n}\mathrm{o}\mathrm{s}\mathrm{i}{\mathrm{s}}_{\mathrm{i}\mathrm{j}}+\sum \mathrm{\beta }\mathrm{k}\mathrm{c}\mathrm{o}\mathrm{v}\mathrm{a}\mathrm{r}\mathrm{i}\mathrm{a}\mathrm{t}\mathrm{e}{\mathrm{s}}_{\mathrm{i}} \end{gathered}$$

Cancer survivors experienced a short-term decline in memory following their diagnoses. For example, in models with years of education mean-centered at 12 years, cancer survivors with 12 years of education experienced a short-term decline in memory of −0.06 SD units (95% CI: −0.08, −0.04, Table 2). This short-term decline was similar in magnitude to having one additional year of typical memory aging at the rate observed between ages 65 and 75 years. However, cancer survivors experienced more favorable long-term memory trajectories before and after their diagnosis than individuals with similar education who did not have a cancer diagnosis; these differences were small in magnitude but statistically significant (Table 2). For example, a cancer survivor with 12 years of education diagnosed with cancer at age 75 experienced, in the subsequent decade, an average rate of memory decline that was 0.058 SD units (95% CI: 0.019, 0.098) slower than those who never had cancer during follow-up. This difference in memory slope corresponded to an average 4.0% (95% CI: 1.27%, 7.15%) slower rate of memory decline for cancer survivors relative to cancer-free individuals with similar educational attainment in the decade following their diagnoses.

Continuous years of education did not modify the cancer-memory decline relationship based on our specified model (Table 2). The short-term decline in memory following a cancer diagnosis was 0.005 SD units (95% CI: −0.031, 0.013, p = 0.22) smaller with each additional year of education after high school (Table 2). The difference between long-term post-diagnosis memory decline and aging-related memory decline was 0.002 SD units per decade slower (95% CI: −0.010, 0.014, p = 0.71) with each additional year of education after high school (Table 2). Sensitivity analyses that excluded cancer survivors with no post-diagnosis memory assessments gave similar results (Supplemental Table 3).

3.2. Effect Modification of Short-Term and Long-Term Post-Diagnosis Decline in Memory by Educational Attainment Categories

Among participants who did not receive a cancer diagnosis during follow-up, those who completed high (12 to <16 years of school) or intermediate (≥16 years) educational level experienced slower aging-related memory decline than participants with low educational level (< 12 years) (Figure 2, Table 3). Among this cancer-free group, in the decade between age 65 and 75 years, the average aging-related memory decline was 0.93 SD units (95% CI: −0.92, −0.94) for those with low education (<12 years), 0.77 SD units (95% CI: −0.78, −0.76) for those with intermediate education (12 to <16 years), and 0.67 SD units (95% CI: −0.68, −0.66) for those with high education (≥16 years). In the decade between age 75 and 85 years, the average aging-related memory decline was 1.49 SD units (95% CI: −1.53, −1.46) for those with low education, 1.37 SD units (95% CI: −1.39, −1.35) for those with intermediate education, and 1.30 SD units (95% CI: 1.33, −1.26) for those with high education. Estimated coefficients used to calculate these outcomes are presented in Table 3.

Figure 2. Memory Score Trajectories from Linear Mixed-Effect Models by Education Categories.

The “Pre-Cancer” and “Post-Cancer” curves represent predicted memory scores for a hypothetical person diagnosed with cancer at age 75 years who has covariate values in their reference categories, across the three levels of education (12 or fewer years, 12 to <16 years, and 16 or more years). The vertical line at age 75 indicates the age at cancer diagnosis. The “Pre-Cancer” curve to the left of cancer diagnosis (time to cancer diagnosis) indicates the long-term change in memory score before cancer diagnosis, and the “Post-Cancer” curve to the right of cancer diagnosis (time since cancer diagnosis) indicates the long-term change in memory score after cancer diagnosis. The discontinuity in the trajectory at cancer diagnosis at age 75 indicates the mean short-term decrement in memory score following diagnosis. The “No Cancer” curves represent predicted memory scores for a hypothetical person who remained cancer-free over the follow-up, also with covariate values in the reference categories, and mean age centered at 75 years. The shaded areas surrounding the curves represent the 95% confidence intervals.

Table 3.

Estimated Regression Coefficients from Linear Mixed Models for Memory Function and Annual Rate of Memory Change by Education, US Health and Retirement Study, United States, 1998–2016

Characteristic	Educational attainment categories
	Model 2: <12 yearsa		Model 3: 12 to <16 yearsa		Model 4: ≥16 yearsa
	(n = 3 769)		(n = 7 918)		(n = 2 762)
	β (95% CI)	Interact ion P-valueb	β (95% CI)	Interact ion P-valueb	β (95% CI)
Memory function (SD units) and memory change (SD units/decade)
Participants with no cancer during follow-up:
Memory function at age 75	−0.989 (−1.063, −0.916)	<0.001	−0.822 (−0.853, −0.792)	<0.001	−0.762 (−0.867, −0.656)
Memory change with linear age	−1.209 (−1.232, −1.187)	<0.001	−1.070 (−1.084, −1.056)	<0.001	−0.983 (−1.008, −0.959)
Memory change with quadratic age, SD units/decade2	−0.281 (−0.293, −0.270)		−0.300 (−0.306, −0.294)		−0.313 (−0.323, −0.303)
Participants with an incident cancer diagnosis:
Difference in memory score right before cancerc	0.098 (0.034, 0.161)	0.637	0.070 (0.033, 0.107)	0.137	−0.016 (−0.075, 0.042)
Change in memory at the time of diagnosis	−0.100 (−0.153, −0.048)	0.154	−0.052 (−0.084, −0.021)	0.622	−0.038 (−0.084, 0.008)
Difference in memory slope before diagnosisc	0.053 (−0.011, 0.116)	0.400	0.062 (0.028, 0.095)	0.073	−0.017 (−0.071, 0.036)
Difference in memory slope after diagnosisc	0.051 (−0.041, 0.142)	0.981	0.050 (−0.002, 0.101)	0.252	0.089 (0.015, 0.163)

^aModels adjusted for sex, race, Southern birthplace, childhood socioeconomic index, total household wealth, height, body mass index, vigorous physical activity, current and ever smoking, alcohol use, childhood self-rated health, and comorbidities (hypertension, diabetes, heart disease, stroke, lung disease)

^bP-values are from the model including interactions terms with each educational attainment category, with the reference group ≥16 years of education.

^cCompared to participants with no cancer over the follow-up, as the reference group

Cancer survivors with high education had similar memory function and slopes before their diagnosis relative to participants with no cancer diagnosis of the same age and educational level (Figure 2, Table 3). In contrast, cancer survivors with low and intermediate education had more favorable memory function and slopes before diagnosis than cancer-free people of the same age and educational level (Figure 2, Table 3). The average age-adjusted difference in memory function for cancer survivors prior to their diagnosis relative to cancer-free individuals was 0.082 SD units (95% CI: 0.018, 0.146) for those with low education and 0.066 SD units (95% CI: 0.029, 0.102) for those with intermediate education (Figure 2, Table 3).

The magnitude of short-term decline in memory was higher in cancer survivors with low education (−0.098 SD units, 95% CI: −0.150, −0.045) than in cancer survivors with high education (−0.036 SD units, 95% CI: −0.082, −0.010), but the interaction with education was not statistically significant (p=0.15). Educational attainment did not modify the difference in long-term memory change experienced by cancer survivors following their diagnosis, as compared to long-term aging-related memory change experienced by cancer-free individuals (Figure 1, Table 3).

Sensitivity analyses excluding participants with cancer who did not have post-diagnosis memory assessments showed similar estimates for the effect modification of short-term and long-term post-diagnosis memory trajectories by educational attainment (Supplemental Table 2).

4. Discussion

In this population-based study of middle-aged and older US adults followed for up to 18 years, survivors of a first incident cancer experienced a short-term decline in memory after diagnosis that exceeded the expected aging-related memory decline in population controls. Among cancer survivors, average rates of long-term memory decline post-diagnosis were largely consistent with rates of pre-diagnosis memory decline, consistent with the phase shift hypothesis of cancer-related cognitive decline.^43,60 Consistent with previous evidence, education was associated with better memory function and slower memory decline regardless of cancer diagnosis status. Our results suggest that cancer survivors with lower educational attainment may experience higher magnitudes of short-term decline in memory following diagnosis than those with high education, but the estimates were imprecise and require further investigation. Finally, we also observed that cancer survivors demonstrated more favorable long-term memory trajectories before and after the diagnosis than similarly aged cancer-free population controls. These more favorable long-term memory trajectories in cancer survivors were present regardless of their baseline educational level.

Comparison to existing literature

We hypothesized that education, as a marker of cognitive reserve, may modify risk of cancer-related cognitive decline. We identified two small studies addressing cognitive reserve and cancer-related cognitive decline, both using cohorts of breast cancer survivors with up to 18 months of follow-up after diagnosis.^35,37 One study used a reading comprehension instrument as a marker of cognitive reserve among women with breast cancer aged 60–70 years, and found that receipt of chemotherapy was associated with worse processing speed and verbal ability scores six months after treatment among those with low, but not high, reading comprehension scores.³⁷ The second study observed that breast cancer survivors with higher education had better verbal episodic memory performance one year after chemotherapy than those with low education.³⁵ Our study corroborates previous evidence on the potential effect modification of cancer-related cognitive decline by education and extends it to a population-based sample of survivors of all cancer types excluding non-melanoma skin cancer. Our finding that, among cancer survivors, average rates of long-term memory decline post-diagnosis were largely consistent with rates of pre-diagnosis memory decline, favoring the “phase shift” hypothesis of cancer-related cognitive decline.

Our results indicating that cancer survivors had more favorable long-term memory trajectories than similarly aged cancer-free adults have been reported by the authors in prior reports.⁴⁹ These results are consistent with a large body of epidemiological literature demonstrating an inverse association between cancer and dementia risk.^{49,56,61–64}This association between cancer and long-term memory decline could be explained by common methodological biases, such as confounding or selective survival bias. To address confounding, we adjusted for a comprehensive set of variables that are known or plausibly associated with both cancer risk and memory decline in mid- to later-life, including early-life and mid-life socioeconomic factors, health behaviors at baseline, and prevalent comorbid conditions at baseline. However, residual confounding is possible and could explain memory aging differences between individuals with and without cancer. Alternatively, selective survival bias could contribute to our finding of more favorable long-term memory trajectories in cancer survivors compared to cancer-free adults. However, this population-based sample with long-term follow-up has better potential to overcome selection biases than the short-term follow-ups of most clinic-based studies.^56,65 The composite memory score used in our models included proxy assessments of participants’ memories to retain severely impaired individuals that would be otherwise excluded from the sample. Proxy assessments were more common over time during the study follow-up, and this happened with higher frequency in cancer survivors. Studies using data from the HRS has demonstrated that the inclusion of proxy interviews reduces the amount of bias on cognitive assessments related to attrition.⁶⁶

Limitations and strengths

Our study is limited to memory function as the only cognitive outcome. Prior clinic-based studies have shown that cancer survivors report a variety of cognitive complaints including forgetfulness and difficulties with attention, word finding, and multitasking.⁶⁷ Objective performance-based assessments of cancer survivors have demonstrated changes in attention, working memory, processing speed, learning, and memory.^31,36,67,68 Future studies should investigate non-memory domains of cognitive function that have previously been implicated in cancer-related cognitive decline.³⁷ We were not able to evaluate associations for individuals who identified as Hispanic, as the memory scores used in this analysis did not validate well in this population subgroup.⁵⁰ Additionally, due to small numbers of racial groups, we were not able to include more granular racial categories in our analysis. Future studies should evaluate whether these findings extend to other diverse populations of cancer survivors.

We had limited statistical power to stratify results by type of cancer or treatment modalities received, which would have resulted in complex four-way interactions in our statistical models. Our results thus represent a population average across different cancer types among middle-aged and older US cancer survivors. Cognitive change in adults has been described across different cancer types. Brain tumors were the first cancer type associated with cognitive change due to a direct effect of these cancers and their treatments on the brain.⁶⁷ However, increasing evidence is available on cognitive change after diagnosis of breast, prostate, colon, and hematological cancers.⁶⁷ Given our limited data on individual cancer types, it is unknown whether the observed short-term decline in memory is only experienced by survivors of these common types of cancers or whether it extends to more rare malignancies. Cancer is frequently treated with combination of treatment modalities (i.e., chemotherapy, radiotherapy, surgery, endocrine, or immunotherapy) for varying periods of time. The HRS collects limited data on cancer treatment, which restricted our ability to describe the types of treatment modalities received by cancer survivors in our study sample. Linkage to Medicare claim records enhances the quality of cancer treatment data, at the expense of restricting the sample size and age range. This linkage is ongoing in the HRS data and will allow investigation of the relationship between cancer treatments and subsequent memory aging, an important future direction. Future research should also investigate how education and other cognitive reserve markers may modify cognition for patients according to specific cancer and treatment types, to better enhance the types of interventions that may be developed in the future. While important to study in the context of memory changes in mid- to later-life, education is an imperfect proxy of cognitive reserve.¹³ Future studies testing theories of cognitive reserve should use constructs that map more accurately onto the conceptual definition of cognitive reserve, such as brain structure measures from neuroimaging in concert with functional measures.^27,69

Strengths of the study include its large, population-based sample of middle-aged and older cancer survivors with comparable cancer-free adults, with data collected biennially over an 18-year follow-up period. Most studies of cancer-related cognitive decline use small, highly select samples of research volunteers recruited in clinical settings with short follow-ups; the present study is an important contribution to this growing body of literature that allows the triangulation of evidence across different types of designs and study populations. Our analysis retained severely cognitively impaired participants that were not able to respond to direct cognitive assessments through incorporating proxy assessments of participants’ memory into our composite memory score, which reduces the probability of any attrition bias that could be introduced by their exclusion.⁵⁰ While attrition bias due to study drop-out is still possible, the linear mixed-effects models used in this analysis are relatively robust to attrition as they use all available observations on all individuals. We adjusted for covariates measured prior to cancer diagnosis, avoiding any misclassification or reverse causation bias. We were able to compare memory function and decline of cancer survivors to similarly aged adults without cancer diagnoses, which allowed us to estimate the magnitudes of short-term and long-term decline in memory associated with a cancer diagnosis in mid-to-late life, over and above aging alone.

Conclusions

In this population-based sample of US middle-aged and older adults, we found that cancer survivors experience a short-term decline in memory in the aftermath of cancer diagnosis compared to memory aging in cancer-free population controls. Cancer survivors with lower educational attainment may experience higher magnitudes of short-term decline in memory following diagnosis than those with high education, but the estimates were imprecise and require further investigation. This population-based sample also demonstrated more favorable long-term memory trajectories before and after the diagnosis than similarly aged cancer-free population controls. These more favorable long-term memory trajectories in cancer survivors may be present regardless of their baseline educational level.