Financial and Health Literacy Predict Incident Alzheimer’s Disease Dementia and Pathology

Participants

Data came from an ongoing longitudinal clinical pathologic cohort study of aging and dementia, the Rush Memory and Aging Project (MAP) [13]. The study was approved by the Institutional Review Board of Rush University Medical Center. Participants were enrolled without known dementia, underwent annual clinical evaluations, and agreed to organ donation after death. A written informed consent and an anatomical gift act were obtained from each participant.

MAP started in 1997 and assessment of literacy was introduced in 2010. At the time of this study, 1,903 participants had enrolled and 1,847 had complete baseline evaluation for the parent study. 618 participants had died and 78 had withdrawn before the literacy assessment. We excluded 72 participants who had severe difficulties with language, hearing, vision, understanding, or no longer living in the geographical area. Of the remaining 1,079 eligible participants, 979 (90.7%) participants had completed at least 1 literacy assessment. Primary analysis on incident AD dementia was conducted in 805 participants who were free of dementia at baseline literacy assessment, and had at least 1 follow-up clinical evaluation. The age at analytic baseline on average was 81.5 years (standard deviation [SD]: 7.7, range: 58.8–100.2). 76.7% were females and the mean years of education was 15.3 (SD: 3.0, range: 5–28).

Literacy and cognitive assessment

Literacy was measured annually using a battery of 32 questions [14]. The battery was designed to assess domain specific knowledge of financial and health information relevant to older population. The financial literacy component consists of 23 questions, many of which were adapted from the Health and Retirement Survey. These questions assess knowledge on financial concepts as well as ability to perform simple monetary calculation. The health literacy component consists of 9 questions, which assess knowledge on Medicare, leading causes of death in old age, and the ability to follow prescription instructions and understand drug risk. Each of the 32 questions was scored correct or incorrect, and the percent of total correct was calculated separately for financial and health literacy components. A composite score for total literacy was obtained by averaging the two percentages. We focused on the total literacy measure throughout the analyses because the composite score covers both skills and is statistically more appealing distribution-wise. Further, we observed a strong positive correlation between financial and health literacy components (Pearson’s r: 0.44, p < 0.001). Total literacy has also been related to various health behaviors and outcomes in older persons [2, 15–20].

Annual cognitive assessment was performed using a battery of 19 tests. The assessment covers a broad range of cognitive processes that are commonly affected by aging, including episodic memory (7 tests), semantic memory (3 tests), working memory (3 tests), perceptual speed (4 tests), and visuospatial ability (2 tests). Raw scores from each test were standardized using baseline mean and standard deviation of the entire cohort, and then averaged to obtain a composite measure of global cognition. Higher scores indicate better cognitive performance. Psychometric properties of these cognitive tests and application of the summary score have been extensively validated in this and other cohorts [21–24]. For the purpose of this study, baseline literacy and cognitive measures were used as predictors in relation to incident AD dementia and subsequently neuropathologic indices of AD.

Clinical diagnosis of AD dementia

Annual diagnosis of AD dementia follows standard criteria, and the decision rules implemented in this cohort have been described previously [25]. Briefly, cognitive assessment results were reviewed by neuropsychologists for signs of cognitive impairment, and clinical diagnoses were made by clinicians after reviewing all available data and examination of the participants. The AD diagnosis requires a history of cognitive decline and evidence of impairment in memory and at least one other cognitive domain. Participants who were not cognitively normal but did not meet clinical criteria for dementia were classified as mild cognitive impairment (MCI) [26].

Neuropathologic indices of AD

All participants in the MAP cohort agreed to organ donation after death. Brain autopsy was performed following a standard procedure [27]. Each brain was cut coronally into 1-cm slabs, and slabs designated for diagnosis purpose were fixed in 4% paraformaldehyde. Using modified Bielschowsky silver stain, counts of neuritic plaques, diffuse plaques, and neurofibrillary tangles were quantified using tissues from five predetermined brain regions (midfrontal, midtemporal, inferior parietal, entorhinal cortices, and hippocampus) [28]. For each of the three markers, raw counts from each region were first scaled using region specific standard deviation and then averaged across all the five regions. A composite measure for global AD pathology was obtained by averaging scores for plaques and tangles. Since the measure was right skewed, square root transformation was applied prior to the analysis. For descriptive purpose, a pathologic AD diagnosis was defined by intermediate or high likelihood according to the modified National Institute on Aging (NIA)-Reagan criteria [29].

Statistical analysis

Pearson correlations and student t-tests described bivariate relationship between demographic, cognitive, and literacy measures at baseline. Cox proportional hazards models examined the associations with incident AD dementia. In these models, the outcome variable was time in years to incident AD dementia (event time). Time was right-censored at death or last clinical evaluation for participants with no event. We first fit a model including terms for age, sex, education, and total literacy, and we repeated by adding a term for cognition.

The model performance in predicting incident AD dementia was assessed using the indices for integrated discrimination improvement (IDI) [12] and continuous net reclassification improvement (NRI) [30]. Briefly, let p (Z_(old) ; t₀) and p (Z_(new) ; t₀) be the risks of AD dementia at time t₀ where Z_(old) and Z_(new) represent covariate vectors under two working survival models, and let T be the event time. If Z_(new) predicts better than Z_(old), we expect that p (Z_(new) ; t₀) > p (Z_(old) ; t₀) for event (i.e., T < t₀), and p (Z_(new) ; t₀) < p (Z_(old) ; t₀) for nonevent (i.e., T > t₀). The continuous NRI statistic can be defined by, P { p ( Z ( new ) ; t 0 ) > p ( Z ( old ) ; t 0 ) | T < t 0 } - P { p ( Z ( new ) ; t 0 ) > p ( Z ( old ) ; t 0 ) | T > t 0 } .

Similarly, the IDI statistic can be defined as the difference in mean risks for population with and without events E { p ( Z ( new ) ; t 0 ) - p ( Z ( old ) ; t 0 ) | T < t 0 } - E { p ( Z ( new ) ; t 0 ) - p ( Z ( old ) ; t 0 ) | T > t 0 } .

We elected these two statistics over traditionalc-statistic because recent literatures suggest that the c-statistic is not sensitive enough to evaluate clinically useful markers [31, 32].

Finally, for participants who had died and undergone brain autopsies (n = 173), multivariable linear regression was performed to examine the independent association of total literacy with AD pathology. In these models, square root transformed global AD pathology was the continuous outcome. The covariates of sex and education were not associated with AD pathology in the initial screening; we therefore only controlled for age at death and cognition for model parsimony.

Statistical inference for IDI and NRI was performed using a R package (survIDINRI) [33], and the rest of analyses were programed using SAS/STAT^® software, Version 9.3 of the SAS System for Linux (SAS Institute, Cary, NC).

Association of literacy with incident AD dementia

Participants were followed for an average of 3.7 years (SD: 1.7, range: 1–7 years), during which 102 (12.7%) were diagnosed with AD dementia. As shown in Table 2, in a model adjusted for age, sex, and education, lower literacy was associated with higher risk for incident AD dementia (hazard ratio [HR]: 2.68, 95% confidence interval [CI]: 2.10–3.41, p < 0.001). After further controlling for global cognition, this association, while attenuated, remained significant (HR: 1.50, 95% CI: 1.14–1.99, p = 0.004). Fig. 1 illustrates cumulative hazards of incident AD dementia for representative participants with different levels of literacy. We repeated the analysis by controlling for five specific cognitive domains, each separately. The domain analyses revealed even stronger associations for total literacy (Table 3).

To further assess the robustness of the finding on literacy and incident AD dementia, we conducted a series of secondary analyses. First, we examined the associations of financial and health literacy components with AD dementia separately. After controlling for demographics and global cognition, lower health literacy was associated with higher risk for AD dementia (HR: 1.43, 95% CI: 1.12–1.82, p = 0.004). Financial literacy showed a similar trend, but the association did not reach the traditional threshold for statistical significance (HR: 1.21, 95% CI: 0.92–1.59, p = 0.164).

Next, considering that the association of literacy with incident AD dementia may differ between participants starting with NCI and those with MCI, we repeated our primary Cox proportional hazards model using MCI status as strata and examined whether the literacy association was dependent on MCI status (i.e., strata interaction). The likelihood ratio test showed that the model with terms for strata interactions did not significantly improve the fit (χ² = 3.17, degrees of freedom = 5, p = 0.674); this suggests that the association of literacy with incident AD dementia did not vary by MCI status. Further, in the stratified model without terms for strata interactions, the result for literacy association with incident AD dementia remained essentially the same as in the primary model (HR: 1.53, 95% CI: 1.16–2.03, p = 0.002), suggesting that the association of literacy with AD was robust against diagnostic status.

Improvement of AD prediction due to literacy

To examine whether the literacy measure improved the prediction of AD above and beyond global cognition, we next evaluated the predictive performance of the model with demographics, cognition and literacy in comparison with the model with only demographics and cognition. Fig. 2 illustrates the participant specific risk scores estimated from the two models. Notably, for participants diagnosed with AD dementia (left panel), 64% had higher risk scores under the model with literacy measure than the model without, indicating improved sensitivity. For participants without AD dementia (right panel), 55% had lower risk scores under the model with literacy measure, indicating improved specificity. The graphical comparisons of the estimated risk scores suggest an improvement in predictive performance when the literacy measure was added to the model. Indeed, test statistics showed that the model with total literacy resulted in significant improvement in risk prediction. Specifically, for a 5-year AD risk, the estimated increase in IDI was 0.023 (95% CI: 0.003–0.058, p = 0.018) and the estimated increase in continuous NRI was 0.245 (95% CI: 0.040–0.364, p = 0.030).

Consistent with the findings from the association analysis, the models replacing global cognition with individual domains showed stronger results for improvement of risk prediction due to literacy. The estimated increase in IDI ranged from 0.055 to 0.124 and the estimated increase in continuous NRI ranged from 0.298 to 0.434 (all ps < 0.01).

Association of literacy with AD pathology

To further confirm the independent relationship between literacy and AD, we leveraged neuropathology data in 173 autopsied participants and examined the association of total literacy with AD pathology. Of this group, 46.5% were NCI, 33.5% had MCI, and 20.0% had AD dementia at death. A majority (63.8%) met criteria for a pathologic diagnosis of AD. Baseline total literacy was lower in participants with pathologic AD (Mean: 57.5%, SD: 14.1%) than those without (Mean: 65.0%, SD: 13.2%, p = 0.001). In a multivariable regression model adjusted for age at death and global cognition, lower total literacy was associated with higher burden of global AD pathology (β: 0.07, standard error [SE]: 0.03, p = 0.035). Notably, in this small sample, baseline global cognition explained approximately 7% of the total variance in global AD pathology, and baseline total literacy explained an additional 2% above and beyondcognition.