Abstract Objectives Genetic risks for cognitive decline are not modifiable; however their relative importance compared to modifiable factors is unclear. We used machine learning to evaluate modifiable and genetic risk factors for Alzheimer’s disease (AD), to predict cognitive decline. Methods Health and Retirement Study participants, aged 65–90 years, with DNA and >2 cognitive evaluations, were included (n = 7,142). Predictors included age, body mass index, gender, education, APOE ε4, cardiovascular, hypertension, diabetes, stroke, neighborhood socioeconomic status (NSES), and AD risk genes. Latent class trajectory analyses of cognitive scores determined the form and number of classes. Random Forests (RF) classification investigated predictors of cognitive trajectories. Performance metrics (accuracy, sensitivity, and specificity) were reported. Results Three classes were identified. Discriminating highest from lowest classes produced the best RF performance: accuracy = 78% (1.0%), sensitivity = 75% (1.0%), and specificity = 81% (1.0%). Top ranked predictors were education, age, gender, stroke, NSES, and diabetes, APOE ε4 carrier status, and body mass index (BMI). When discriminating high from medium classes, top predictors were education, age, gender, stroke, diabetes, NSES, and BMI. When discriminating medium from the low classes, education, NSES, age, diabetes, and stroke were top predictors. Discussion The combination of latent trajectories and RF classification techniques suggested that nongenetic factors contribute more to cognitive decline than genetic factors. Education was the most relevant predictor for discrimination. Cognitive decline, Cognitive trajectories, Machine learning, Random forests, Risk factors Risk factors for cognitive decline presumably include many of the same factors that put individuals at risk for Alzheimer’s disease (AD) and dementia and also coincide with known risks for cardiovascular disease. Prior studies have evaluated such modifiable risk factors in the context of population attributable risks (PARs) for AD and dementia to motivate public health efforts to control or reduce the prevalence of risk factors for AD and dementia (Norton, Matthews, Barnes, Yaffe, & Brayne, 2014). Others have developed risk scores incorporating multiple, potentially modifiable, risk factors and assigning weights to provide an easy-to-use guide for clinicians and patients to evaluate their risks (Kaffashian et al., 2013; Kivipelto et al., 2006). Such risk scores are often based on, or are parallel to, well-known cardiovascular risk scores (Kaffashian et al., 2013). These methods for evaluating risks provide useful information for the research community, clinicians, and for individuals. However, PARs and risk scores do not provide an overall evaluation of the relative rankings of these factors. Such a ranking would show which risk factors have the greatest potential to reduce disease risk on a population level. PARs tend to be generated on one factor at a time and based on population prevalence of a condition, making it difficult to evaluate one risk in the presence of another. Risk scores assign weight based on prior studies, not necessarily derived from one nationally representative cohort, and can be formed based on one risk factor at a time. This distinction is important because people are exposed to multiple risk factors at the same time. A ranking of the importance of various risk factors, derived from a multivariate model, can help put risk factors into context from a public health perspective. We use machine learning (ML) methods to infer multivariate prediction models and to determine the relative importance of cognitive decline risk factors from a population-based sample. ML methods are considered to be a branch of the artificial intelligence field and they comprise a set of computational algorithms with a strong connection to statistics. One of the main goals of ML is to make accurate predictions based on data. They have become increasingly popular for their capability to deal with high-dimensional data and uncover complex patterns present in the data. In this work, we use Random Forests (Breiman, 2001) (RF) for classification to evaluate risk factors of cognitive decline in the Health and Retirement Study (HRS) cohort. RF since its inception in 2001 has become a very popular approach in the ML and bioinformatics communities and has motivated a large body of research to understand its behavior and produce improvements (Chen & Ishwaran, 2012; Strobl, Boulesteix, Kneib, Augustin, & Zeileis, 2008). This momentum led to the creation of a whole family of ML methods called Decision Tree Ensembles (DTE) (Brandmaier, Prindle, McArdle, & Lindenberger, 2016; Brick, Koffer, Gerstorf, & Ram, 2017; Chen & Ishwaran, 2012; Strobl, Malley, & Tutz, 2009). RF is highly nonlinear, multivariate, and can deal with high-dimensional data. RF contains built-in metrics of variable importance which allow for the evaluation of the relative relevance of each variable in a RF model. These features make RF methods an appealing tool to evaluate the importance of various risk and protective factors that influence cognitive decline. Such risk factors include demographic characteristics, health characteristics, and sociocultural characteristics including educational attainment and neighborhood socioeconomic status (NSES). There is a substantial literature supporting most of these factors’ associated risks and benefits for cognitive function in aging. However, there is limited data evaluating the effects of NSES on cognitive decline, and few studies that simultaneously evaluate genetic and caridovasulcar risks in the same model. We sought to evauate the relative importance in one model, of each of these factors in addition to SNPs that have been associated with AD as well as SNPs that have been associated with stress and inflammation. Methods We used data from the HRS to evaluate risk factors for cognitive decline. The HRS was assembled by merging cohorts from several similar studies with the same general foci, that were initiated around the same time (i.e., Asset and Health Dynamics among the Oldest Old (AHEAD) study; the War Baby Study; and the Children of the Depression Study). Today, these studies are collectively referred to as the HRS and form a large, longitudinally followed, representative cohort of Americans aged 50 years and older. All participants provided informed consent. Interviews take place biennially and are conducted by the Survey Research Center at the University of Michigan. The study protocol was approved by the University of Michigan Institutional Review Board (IRB). The current project was approved by the Wake Forest School of Medicine and the Duke University Medical Center Institiutional Review Boards. Participants At each participants’ first interview, a cognitive test is administered. Repeated biennial cognitive evaluations only begin once a participant turns age 65. The current study includes only observations from participants who were aged 65 years or older and is further limited to those who participated in DNA collection in 2006 or 2008. DNA HRS genotype data was obtained from dbGAP (www.ncbi.nlm.nih.gov/projects/gap/cgi-bin/study.cgi?study_id=phs000428.v1.p1.). Briefly, saliva samples were collected for DNA extraction and genome-wide association studies (GWAS) in 2006 and 2008. Illumina Human Omni-2.5 Quad bead chips were used for genotyping. Details for the genotyping procedure and the quality control approach that was applied for all of the genotypes used in the analyses in this paper are provided in the Quality Control Report for Genotypic Data for dbGaP users of the HRS genotypic data http://hrsonline.isr.umich.edu/sitedocs/genetics/HRS_QC_REPORT_MAR2012.pdf). We used a subset of SNPs from the HRS subject-level genotype matrix format file. All of the genotypes in this matrix passed QC filters that were defined in the Quality Control Report for Genotypic Data for dbGaP users of the HRS genotypic data. A total of 12,507 study subjects were genotyped on the Illumina HumanOmni2.5-4v1 array. The median call rate is 99.7% and the error rate estimated from 336 pairs of study sample duplicates is 6 × 10–5. SNPs for the current analysis were selected based on meta-analyses of susceptibility genes for AD (APOE, BIN1, CD2AP, CD33, CLU, CR1, MS4A4E, MS4A6A, PICALM, and TOMM40) (Bertram, McQueen, Mullin, Blacker, & Tanzi, 2007; Lambert et al., 2013). A second set of SNPs from 10 genes (BDNF, CRP, IFN-G874T, IL-IBC-511T, IL-6, IL-10, MTHFR, PPARG, SLC6A4, and TNFα) were selected based on their associations with stress, inflammation, and cardiovascular disease, as immune system and inflammatory pathways are influenced by both environmental exposures and some of the same genes that are linked to AD and cognitive decline(Elkins et al., 2007; Krabbe et al., 2009; Marioni et al., 2010; Mooijaart et al., 2011; Payton et al., 2005; Sanchez et al., 2011; Tsai et al., 2010; Yaffe et al., 2008). Selected SNPs were weighted based on International Genomics of Alzheimer’s Project (IGAP) beta weights (Lambert et al., 2013). To prevent spurious association due to population stratification, we selected only non-Hispanic Caucasian participants as identified in the data set. Cognitive Function The HRS uses an abbreviated version of the modified Telephone Interview for Cognitive Status (TICS) to evaluate cognitive function (Brandt, Spencer, & Folstein, 1988; Welsh, Breitner, & Habib-Magruder, 1993). The TICS is based on the Mini-Mental State Examination (Folstein, Folstein, & McHugh, 1975) and is an established instrument for the evaluation of global cognitive function. For the HRS, the TICS was modified to an abbreviated version totaling 35 points. Neighborhood Socioeconomic Status NSES was generated by the RAND Corporation and is based on consideration of six neighborhood characteristics including: the percent of adults aged 25 years or older without a high school education; percent male unemployment; percent of households with income below the poverty line; percent of households receiving public assistance; percent of households with a female as head-of-household; and median household income(RAND (2010)Neighborhood SES Index Data Core User’s Documentation Series 2010, 15, 2010). The index (0–100) applied to each participant is based on a link to the census tract of participants’ home residence in 2000. Cardiovascular Risk Factors and Cerebrovascular Disease Participants were asked in their biennial evaluations whether they had hypertension, diabetes, or any history of heart disease, or stroke. To apply the most rigorous characterization of hypertension and diabetes, we required participants to not only indicate that they were diagnosed by a doctor, but that they were currently taking medications for that condition. If participants indicated that they had been told by a doctor that they had any of the following: heart attack, coronary heart disease, angina, congestive heart failure, or other heart problems, they were counted as having a heart condition. If participants indicated that they had been seen by a doctor for a stroke, they were counted as having had a stroke. For each of these conditions, we then coded participants as having the condition ever versus never, as the random forest analysis was unable to accommodate time varying covariates. Other covariates we included were age, gender, and education level. Statistical Approach We combined latent trajectories approaches with ML analyses based on RF in this evaluation. Initially we evaluated participants’ cognitive trajectories using TICS scores. We estimated latent trajectory classes of cognition by first determining the number of trajectory groups using a group-based trajectory model (using PROC TRAJ in SAS). The Bayesian Information Criterion (BIC) was used to select the optimal number of trajectory groups, i.e., the number of trajectory groups was varied until the best-fitting model was obtained as indicated by the BIC. A priori, we determined that smallest trajectory class should include more than 10% of the subjects from the sample in order to provide a clinically representative pattern of change. Intercept only, linear, quadratic, or cubic polynomial terms were modeled. Posterior probabilities of group membership from each individual were used to assess model fit. High probability of membership in a single group represents a good model fit. In order to derive the trajectories of cognitive performance without adjustment, we included only the TICS scores in the model. Demographic and other characteristics were compared across trajectory classes. Continuous variables were evaluated using generalized linear models and categorical variables were evaluated with χ2 statistics. After the trajectory classes were generated, we used RF to evaluate discrimination of these groups based on clinical, demographic and genetic data described above. RF for classification is one of the so-called ensemble methods for classification, because a set of classifiers (instead of one) is generated and each one casts a vote to define the outcome of the model. Each classifier is a tree built using the classification and regression trees methodology (CART)(Breiman, Friedman, Olsen, & Stone, 1984). In problems with relatively few variables, RF requires minimal tuning of the parameters and the default values often produce good results. Once the forest is built, predicting class membership of a new sample is accomplished by combining the trees, using a majority vote. As a result of using a bootstrap sampling of the training data, around one-third of the samples are omitted when building each tree. These are so-called out-of-the-bag (OOB) samples, which can be used to assess the performance of the classifier and to build measures of importance. This OOB mechanism is similar to a built-in cross-validation procedure that allows evaluation of performance. In the present report, we used the permutation index of variable importance which quantifies decreases in accuracy of the estimated RF model due to random permutation of a given variable. Class discrimination was performed in a pair-wise manner while ranks of variables based on permutation index were generated to assess the weight of each variable in the prediction model. We gave preference to RF feature importance measures because RF they can capture complex multivariate and nonlinear relationships between the variables that can be missed by logistic regression beta weights. In our analyses for variable selection, we used a strategy proposed by Strobl and colleagues for RF (Strobl et al., 2009). They suggested discarding as noisy, the variables with negative permutation index and also the positive ones with absolute values less than the amplitude of the negative score with maximum amplitude. This strategy was used recently by Dermott and colleagues to investigate memory resilience to AD risk (McDermott, McFall, Andrews, Anstey, & Dixon, 2017b) and by Kaup and colleagues to study cognitive resilience to APOE-e4 (Kaup et al., 2015). To evaluate performance in each case, data were subsampled to generate 10 balanced datasets with sample sizes 2,000, 2,800, and 2,000, respectively. Average and standard deviation values of OOB estimates of accuracy, sensitivity and specificity were reported. Average ranks of variables were also generated. Statistical software, SAS 9.4 (SAS Institute, Inc.) was used for the derivation of latent trajectory classes and other comparisons among classes. RF analysis was conducted using the randomForest R library(Liaw & Wiener, 2002) with ntree = 500 and mtry = 20. We selected mtry = 20 to alleaviate problems of the permutation index in the presence of correlated predictors (Strobl et al., 2008). Results Of a total of 12,507 participants with DNA, 8,709 were aged 65 years or older, had follow-up evaluations, had complete data for genes of interest, and were not missing TICS scores, or education level. We set aside 1,258 non-White participants for this analysis and 309 participants who had missing APOE or NSES information. The latent trajectory classes analysis using PROC TRAJ revealed three classes including a group (high performing) that had the highest intercept and remained fairly high functioning over time with a slight decline. There was a middle performing group with an intermediate level intercept and moderate decline over time, and the lowest performance group (low) had the lowest intercept and a slightly steeper level of decline toward the end of the follow-up period (Figure 1). For this study, the mean number of observations per participant is 6.13 [SD = 2.41] with a range of 2–10. We have included Supplementary Table S1 (in Supplementary Materials) of parameter estimates for the intercepts, linear and quadratic growth and their standard errors for each of the classes along with T-statistics. A test of equality of intercepts, linear and quadratic terms across the classes reveal that there is significant difference in the intercept and quadratic terms between the classes as well as on their pairwise comparisons. The intercepts are significantly different from each other (p-values of the chi-square-test for various pairwise comparisons of intercepts are all p-value < .0001). The quadratic slope comparison between “Low” and “Medium” groups is not significant (p-value = .9543), while there is significant difference in the quadratic slopes between “Medium” and “High” (p-value < .0001) and the groups “Low” and “High” group (p-value < .0015). The variability of these growth parameters within each class is rather minimal. We have included in supplementary materials a plot of BIC versus number of classes (Supplementary Figure S1). The final number of classes for TICS is determined by the highest BIC value combined with percent of observations in any class being greater than 10% of the sample. We have also included a table of predicted values of the TICS along 95% confidence interval used by PROC TRAJ to plot the trajectory classes (Supplementary Table S2). Figure 1. View largeDownload slide The latent trajectory classes analysis using PROC TRAJ revealed three classes. The predicted values and 95% confidence intervals are plotted for each class. Figure 1. View largeDownload slide The latent trajectory classes analysis using PROC TRAJ revealed three classes. The predicted values and 95% confidence intervals are plotted for each class. Comparisons of demographic risk factors across these groups revealed that higher performing participants were younger, more highly educated, less likely to carry an APOE ε4 allele and healthier in general (e.g., diabetes, stroke, heart disease, etc.; Table 1). Table 1. Demographic Characteristics of 7,142 HRS Participant by Latent Trajectory Class Characteristics 1 Low Trajectory Class 2 Medium Trajectory Class 3 High Trajectory Class Total n 1,012 3,319 2,811 7,142 p-value Age (SD) 68.6 (4.5) 67.7 (3.5) 67.0 (2.8) 67.6 (3.5) <.0001 Female Gender (%) 531 (52.5) 1,762 (53.1) 1,787 (63.6) 4,080 (57.1) <.0001 Education (years; SD) 9.5 (3.8) 12.4 (2.6) 13.9 (2.3) 12.6 (3.1) <.0001 APOE ε4+ (%) 320 (31.6) 868 (26.2) 638 (22.7) 1,826 (25.6) <.0001 BMI (SD) 27.9 (5.3) 27.6 (5.2) 27.0 (5.0) 27.4 (5.1) <.0001 Hypertension (%) 751 (74.2) 2,246 (67.7) 1,858 (66.1) 4,855 (68.0) <.0001 Diabetes (%) 325 (32.1) 718 (21.6) 467 (16.6) 1,510 (21.1) <.0001 Stroke (%) 160 (15.8) 359 (10.8) 184 (6.5) 703 (9.8) <.0001 Any Heart (%) 458 (45.3) 1,435 (43.2) 1,058 (37.6) 2,951 (41.3) <.0001 NSES (SD) 74.8 (8.3) 78.1 (6.2) 79.6 (5.8) 78.3 (6.6) <.0001 Characteristics 1 Low Trajectory Class 2 Medium Trajectory Class 3 High Trajectory Class Total n 1,012 3,319 2,811 7,142 p-value Age (SD) 68.6 (4.5) 67.7 (3.5) 67.0 (2.8) 67.6 (3.5) <.0001 Female Gender (%) 531 (52.5) 1,762 (53.1) 1,787 (63.6) 4,080 (57.1) <.0001 Education (years; SD) 9.5 (3.8) 12.4 (2.6) 13.9 (2.3) 12.6 (3.1) <.0001 APOE ε4+ (%) 320 (31.6) 868 (26.2) 638 (22.7) 1,826 (25.6) <.0001 BMI (SD) 27.9 (5.3) 27.6 (5.2) 27.0 (5.0) 27.4 (5.1) <.0001 Hypertension (%) 751 (74.2) 2,246 (67.7) 1,858 (66.1) 4,855 (68.0) <.0001 Diabetes (%) 325 (32.1) 718 (21.6) 467 (16.6) 1,510 (21.1) <.0001 Stroke (%) 160 (15.8) 359 (10.8) 184 (6.5) 703 (9.8) <.0001 Any Heart (%) 458 (45.3) 1,435 (43.2) 1,058 (37.6) 2,951 (41.3) <.0001 NSES (SD) 74.8 (8.3) 78.1 (6.2) 79.6 (5.8) 78.3 (6.6) <.0001 Note: BMI = Body mass index; HRS = Health and Retirement Study; NSES = Neighborhood socioeconomic status; SD = Standard deviation. View Large Table 1. Demographic Characteristics of 7,142 HRS Participant by Latent Trajectory Class Characteristics 1 Low Trajectory Class 2 Medium Trajectory Class 3 High Trajectory Class Total n 1,012 3,319 2,811 7,142 p-value Age (SD) 68.6 (4.5) 67.7 (3.5) 67.0 (2.8) 67.6 (3.5) <.0001 Female Gender (%) 531 (52.5) 1,762 (53.1) 1,787 (63.6) 4,080 (57.1) <.0001 Education (years; SD) 9.5 (3.8) 12.4 (2.6) 13.9 (2.3) 12.6 (3.1) <.0001 APOE ε4+ (%) 320 (31.6) 868 (26.2) 638 (22.7) 1,826 (25.6) <.0001 BMI (SD) 27.9 (5.3) 27.6 (5.2) 27.0 (5.0) 27.4 (5.1) <.0001 Hypertension (%) 751 (74.2) 2,246 (67.7) 1,858 (66.1) 4,855 (68.0) <.0001 Diabetes (%) 325 (32.1) 718 (21.6) 467 (16.6) 1,510 (21.1) <.0001 Stroke (%) 160 (15.8) 359 (10.8) 184 (6.5) 703 (9.8) <.0001 Any Heart (%) 458 (45.3) 1,435 (43.2) 1,058 (37.6) 2,951 (41.3) <.0001 NSES (SD) 74.8 (8.3) 78.1 (6.2) 79.6 (5.8) 78.3 (6.6) <.0001 Characteristics 1 Low Trajectory Class 2 Medium Trajectory Class 3 High Trajectory Class Total n 1,012 3,319 2,811 7,142 p-value Age (SD) 68.6 (4.5) 67.7 (3.5) 67.0 (2.8) 67.6 (3.5) <.0001 Female Gender (%) 531 (52.5) 1,762 (53.1) 1,787 (63.6) 4,080 (57.1) <.0001 Education (years; SD) 9.5 (3.8) 12.4 (2.6) 13.9 (2.3) 12.6 (3.1) <.0001 APOE ε4+ (%) 320 (31.6) 868 (26.2) 638 (22.7) 1,826 (25.6) <.0001 BMI (SD) 27.9 (5.3) 27.6 (5.2) 27.0 (5.0) 27.4 (5.1) <.0001 Hypertension (%) 751 (74.2) 2,246 (67.7) 1,858 (66.1) 4,855 (68.0) <.0001 Diabetes (%) 325 (32.1) 718 (21.6) 467 (16.6) 1,510 (21.1) <.0001 Stroke (%) 160 (15.8) 359 (10.8) 184 (6.5) 703 (9.8) <.0001 Any Heart (%) 458 (45.3) 1,435 (43.2) 1,058 (37.6) 2,951 (41.3) <.0001 NSES (SD) 74.8 (8.3) 78.1 (6.2) 79.6 (5.8) 78.3 (6.6) <.0001 Note: BMI = Body mass index; HRS = Health and Retirement Study; NSES = Neighborhood socioeconomic status; SD = Standard deviation. View Large Best RF performance was achieved when discriminating high from low trajectory classes. RF produced accuracy = 78% (0.4%), sensitivity = 75% (1.0%), and specificity = 81% (1.0%). The top ranked predictors in this comparison were education, age, gender, stroke, NSES, diabetes, APOE carrier status, and body mass index (BMI) (Figure 2). When discriminating the high from medium cognitive performance classes RF produced accuracy = 63% (0.2%), sensitivity = 61% (1.0%), and specificity = 64% (0.3%). The top ranked predictors in this case were education, age, gender, stroke, diabetes, NSES, and BMI (Figure 3). When discriminating the medium from the low trajectory classes, RF produced accuracy = 66% (1.0%), sensitivity = 61% (1.0%), and specificity = 71% (1.0%). The top ranked predictors in this case were education, NSES, age, diabetes, and stroke (Figure 4). To further investigate the high relevance of education in these results, we repeated these analyses by excluding the education in the RF model (not presented). Excluding education from the model led to large decreases of the models’ accuracy (around 10%). Figure 2. View largeDownload slide Relative variable importance when discriminating high performing versus low performing groups. BMI = Body mass index; SES = Socioeconomic status Figure 2. View largeDownload slide Relative variable importance when discriminating high performing versus low performing groups. BMI = Body mass index; SES = Socioeconomic status Figure 3. View largeDownload slide Relative variable importance when discriminating high performing versus medium performing groups. BMI = Body mass index; SES = Socioeconomic status Figure 3. View largeDownload slide Relative variable importance when discriminating high performing versus medium performing groups. BMI = Body mass index; SES = Socioeconomic status Figure 4. View largeDownload slide Relative variable importance when discriminating medium performing versus low performing groups. BMI = Body mass index; SES = Socioeconomic status. Figure 4. View largeDownload slide Relative variable importance when discriminating medium performing versus low performing groups. BMI = Body mass index; SES = Socioeconomic status. Discussion Our RF analysis of modifiable and nonmodifiable predictors of the latent trajectory classes of cognitive decline revealed that the most prominent predictor of class membership was education level. The sensitivity and specificity of the discrimination was best for the highest performing versus lowest performing group. Accuracy, sensitivity, and specificity for other comparisons were lower. Nonetheless, the ranking of the factors across groups was relatively consistent. Education was followed by age, NSES, gender, diabetes, and stroke. These findings suggest that modifiable factors (e.g., education level, NSES, and diabetes) compared to genetic risk factors are more predictive of cognitive decline in the HRS cohort. Higher education levels have been a consistent predictor of cognitive performance across studies of cognitive function over time (Beydoun et al., 2014; Stern et al., 1994; Vemuri et al., 2014), whereas lower education has been a consistent risk factor for poor cognitive performance (Beydoun et al., 2014; Schmand et al., 1997). A recent report with a similar analytic design, i.e., growth mixture models to define classes and random forest analysis to rank predictors, reported that education was a similarly a very strong predictor of cognitive performance (McDermott et al., 2017a). Aside from the obvious association between education level and testing experience, education has been seen as a proxy for cognitive reserve (Stern et al., 1994). Indeed, MRI studies have shown associations between education level and hippocampal volume (Shpanskaya et al., 2014). Higher education levels may potentially buffer individuals from symptoms of neurodegenerative disease, thereby delaying time to diagnosis (Stern, 2012). Education levels have been increasing in the United States since the data were first collected in 1940. Since that time, the percentage of Americans with a bachelor’s degree or higher has risen from 4.6% to 33% (Ryan & Bauman, 2016). Recent reports suggest that this may influence the incidence of cognitive impairment, potentially having the effect of reduced prevalence of cognitive impairment and dementia in the coming decades (Norton et al., 2014). Higher levels of education have a positive effect on cognitive performance, mainly reflected as the intercept, or baseline level of cognitive performance in many studies.(Schneider et al., 2012; Wilson et al., 2009; Zahodne et al., 2011) Yet increased population levels of education may be countered by increasing life span (age is the strongest, nonmodifiable risk factor for cognitive decline and dementia), as well as an epidemic of type 2 diabetes and obesity, both risks for cognitive decline and dementia. Nonetheless, the consistently positive associations between education and cognitive performance invite further study into the mechanisms by which education or cognitive reserve serves to maintain cognitive function. Low NSES has been consistently associated with poorer cognitive function in the few studies that have included this and similar variables that seek to capture factors that describe an individual’s immediate environment (Lang et al., 2008; Shih et al., 2011; Zeki Al Hazzouri et al., 2011). The NSES variable used here was developed by the RAND Corporation and represents neighborhood indicators including education, unemployment, poverty, public assistance, female head-of-households, and median income. These factors potentially capture the level of deprivation or affluence in a person’s area of residence and may further suggest the amenities available to residents, or lack thereof. Our understanding of precisely how the local environment influences cognition is limited and perhaps further research may elucidate the most meaningful and readily modifiable factors on individual and population levels. Age and gender are not modifiable risk factors but both were relevant in this analysis. A discussion of age as a risk factor for cognitive decline is beyond the scope of this project as it is a complex issue with a separate body of literature focused on the nature of age-related cognitive decline. The next most important indicator we found was gender. In this study, women performed better than men on the TICS. This is consistent with other studies that have reported gender differences in cognitive performance favoring females on cognitive status and verbal measures (McCarrey, An, Kitner-Triolo, Ferrucci, & Resnick, 2016; Proust-Lima, Amieva, Dartigues, & Jacqmin-Gadda, 2007). Diabetes and stroke were included among the more important potentially modifiable risk factors to discriminate between groups. Both have been associated with cognitive impairment and dementia (Dik et al., 2000; Jefferson et al., 2015; Knopman et al., 2001; Vermeer et al., 2003). Both risk factors have relatively low prevalence in older adults (diabetes 5% and stroke 0.8% in our sample). For this reason, they may register lower PARs and yet clearly have a strong impact on cognitive function at the individual level. In a recent projection of the impact of risk factor reduction, diabetes had an estimated PAR of 3% in the United States; stroke was not ranked. Neither were included in the CAIDE Dementia Risk Score (Kivipelto et al., 2006). However, in a study using data from the National Alzheimer’s Coordinating Centers, higher Framingham Stroke Risk Scores (which included diabetes) were associated with cognitive decline (Jefferson et al., 2015). Finally, APOE ε4 carrier status was the top ranked genetic variable in the discrimination between the highest and lowest performing groups but it was not as strong as one might have expected, and was even less relevant in the other two comparisons. There are several potential reasons for this finding. First, most studies of cognitive decline that investigate the effects of APOE ε4 carrier status report results stratified by APOE (Cosentino et al., 2008). In our study, APOE ε4 carriers were well distributed across groups that were defined by their cognitive performance rather than genetic status (Table 1). Second, a number of studies have investigated the association between APOE ε4 carrier status and rate of decline. Although most indicate that APOE ε4 increases the risk of AD, there is disagreement about whether or not APOE influences the rate of cognitive decline (Boyle, Buchman, Wilson, Kelly, & Bennett, 2010; Bunce, Fratiglioni, Small, Winblad, & Backman, 2004; Christensen et al., 2008; Kleiman et al., 2006). Other genes included in our analysis were similarly less relevant for prediction than modifiable factors. This is not unexpected as our prior work showed that a genetic risk score (summing risks of the top 10 AlzGene (Bertram et al., 2007) AD risk genes) had little influence on cognitive decline in this cohort (Hayden, Lutz, Kuchibhatla, Germain, & Plassman, 2015). A recent GWAS of cognitive decline identified some previously unrecognized variants (Sherva et al., 2014) so it is possible that genes other than those recognized for AD risk contribute to cognitive decline. Finally, some technical considerations about our approach. We have used in this work a two-step approach that combines latent trajectories methods with RF analyses to investigate predictors of cognitive decline. Recently, McDermott and colleagues have proposed another two-step method which combines growth mixture models with a different DTE approach to investigate memory resilience to AD risk (McDermott et al., 2017b). It is worth mentioning that a recent adition to the class of DTE methods, structural equation model forests (Brandmaier et al., 2016) embodies a one step approach that potentially can avoid possible accumulation of errors present in two step approaches. Also since the number of variables (p = 31) was relatively small in our problem and our goal was to assess the relative weight of all these predictors we did not use a feature selection technique as for example recursive feature elimination (Brick et al., 2017). In our analyses we used for variable selection a rule proposed by Strobl and colleagues for RF. But still our results should be interpreted with caution due to potential sources of bias that have been associated to RF permutation index such as correlated predictors and preference for variables with more categories. The latter in our case is alleviated by the fact that we build ranks based on averages of the scores rather than single trial. Although in general correlation between variables were small a few (e.g., Education and NSES r = .38) could be a source of concern. This issue was addressed here by choising the mtry parameter much larger (mtry = 20) than the default in the package (sqrt(31)) (Strobl et al., 2008). Limitations There are several limitations to note. Latent trajectories were estimated using PROC TRAJ which has limitations as it does not incorporate all the model fit statistics available in the literature and does not output all the elements of fit statistics. We only had access to the NSES index at one time point and it is possible that participants did not spend the majority of their lives in the location used to code their NSES level or in a location with a similar NSES level. Although a proportion of participants likely remained in the same residence for many years, it is unclear at what point in a person’s life, a neighborhood exposure may affect cognitive outcomes. Also, the NSES associated with a particular location can change over time, although a recent report on derivation of a time-invariant measure of NSES suggests that NSES can remain relatively stable over time (Miles et al., 2016). Cardiovascular health in this study is a self-reported measure. Participants’ self-report of some cardiovascular conditions were only used if the participant reported that the condition was treated and that a doctor diagnosed the condition in an attempt to obtain a more rigorous indicator of cardiovascular (CVD) status. Due to our inability to account for time in the random forest models, we derived cognitive trajectory groups and classified participants based on their cognitive trajectories over time. To address incidence of CVD or CVD risk factors (hypertension and diabetes), we evaluated self-reported CVD, hypertension, diabetes, and stroke over the course of follow-up and classified participants as ever/never endorsing conditions. Ever reporting the condition may be a better characterization of the risk of cognitive decline to account for risk conveyed by both subclinical and clinical disease. Finally, although PROC TRAJ is a valid and useful approach, it does not estimate class specific variances in intercept and slope, which might be important here. Growth Mixture Modeling in Mplus allows this, as well as provided statistical tests for changes in model fit. Conclusions The combined application of latent trajectories and RF classification techniques suggests that nongenetic factors provided stronger contributions to cognitive decline than genetic factors in this nationally representative cohort. Education, age, NSES, gender, stroke, and diabetes were the most important predictors to discriminate the classes of cognitive decline. Our findings of the importance of modifiable factors are in line with other similar reports(Baumgart et al., 2015; Jefferson et al., 2015; Norton et al., 2014), although our ranking provides additional insight. There is growing consensus on the importance of modifiable risk factors as these are factors that can be addressed now on individual and societal levels to impact cognitive function in aging. Supplementary Material Supplementary data is available at The Journals of Gerontology, Series B: Psychological Sciences and Social Sciences online. Funding This research was supported by the National Institute on Aging, grant # R01AG042633. Conflict of Interest M. Kuchibhatla is a statistical consultant for Scion NeuroStim, LLC. M. W. Lutz receives consulting fees from Zinfandel Pharmaceutics and Cabernet Pharmaceutics. References Baumgart, M., Snyder, H. M., Carrillo, M. C., Fazio, S., Kim, H., & Johns, H. ( 2015). Summary of the evidence on modifiable risk factors for cognitive decline and dementia: A population-based perspective. Alzheimer’s & Dementia , 11, 718– 726. doi: 10.1016/j.jalz.2015.05.016 Google Scholar CrossRef Search ADS Bertram, L., McQueen, M. B., Mullin, K., Blacker, D., & Tanzi, R. E. ( 2007). Systematic meta-analyses of Alzheimer disease genetic association studies: The AlzGene database. Nature Genetics , 39, 17– 23. doi: 10.1038/ng1934 Google Scholar CrossRef Search ADS Beydoun, M. A., Beydoun, H. A., Gamaldo, A. A., Teel, A., Zonderman, A. B., & Wang, Y. ( 2014). Epidemiologic studies of modifiable factors associated with cognition and dementia: Systematic review and meta-analysis. BMC Public Health , 14, 643. doi: 10.1186/1471-2458-14-643 Google Scholar CrossRef Search ADS Boyle, P. A., Buchman, A. S., Wilson, R. S., Kelly, J. F., & Bennett, D. A. ( 2010). The APOE epsilon4 allele is associated with incident mild cognitive impairment among community-dwelling older persons. Neuroepidemiology , 34, 43– 49. doi: 10.1159/000256662 Google Scholar CrossRef Search ADS Brandmaier, A. M., Prindle, J. J., McArdle, J. J., & Lindenberger, U. ( 2016). Theory-guided exploration with structural equation model forests. Psychological Methods , 21, 566– 582. doi: 10.1037/met0000090 Google Scholar CrossRef Search ADS Brandt, J., Spencer, M., & Folstein, M. ( 1988). The telephone interview for cognitive status. Neuropsychiatry, Neuropsychology, and Behavioral Neurology , 1, 111– 117. Retrieved from: https://www.researchgate.net/profile/Jason_Brandt2/publication/232211604_The_Telephone_Interview_for_Cognitive_Status/links/00b7d534486f0b3489000000.pdf Breiman, L. ( 2001). Random forests. Machine Learning , 45, 5– 32. Google Scholar CrossRef Search ADS Breiman, L., Friedman, J. H., Olsen, R. A., & Stone, C. J. ( 1984). Classification and regression trees . Boca Raton: Chapman & Hall/CRC. Brick, T. R., Koffer, R. E., Gerstorf, D., & Ram, N. ( 2017). Feature selection methods for optimal design of studies for developmental inquiry. Journals of Gerontology, Series B: Psychological Sciences and Social Sciences , 73, 113–123. doi: 10.1093/geronb/gbx008 Bunce, D., Fratiglioni, L., Small, B. J., Winblad, B., & Bäckman, L. ( 2004). APOE and cognitive decline in preclinical Alzheimer disease and non-demented aging. Neurology , 63, 816– 821. doi:10.1212/01.WNL.0000137041.86153.42 Google Scholar CrossRef Search ADS Chen, X., & Ishwaran, H. ( 2012). Random forests for genomic data analysis. Genomics , 99, 323– 329. doi: 10.1016/j.ygeno.2012.04.003 Google Scholar CrossRef Search ADS Christensen, H., Batterham, P. J., Mackinnon, A. J., Jorm, A. F., Mack, H. A., Mather, K. A., … Easteal, S. ( 2008). The association of APOE genotype and cognitive decline in interaction with risk factors in a 65-69 year old community sample. BMC Geriatrics , 8, 14. doi: 10.1186/1471-2318-8-14 Google Scholar CrossRef Search ADS Cosentino, S., Scarmeas, N., Helzner, E., Glymour, M. M., Brandt, J., Albert, M., … Stern, Y. ( 2008). APOE epsilon 4 allele predicts faster cognitive decline in mild Alzheimer disease. Neurology , 70, 1842– 1849. doi: 10.1212/01.wnl.0000304038.37421.cc Google Scholar CrossRef Search ADS Dik, M. G., Deeg, D. J., Bouter, L. M., Corder, E. H., Kok, A., & Jonker, C. ( 2000). Stroke and apolipoprotein E epsilon4 are independent risk factors for cognitive decline: A population-based study. Stroke , 31, 2431– 2436. doi:10.1161/01.STR.31.10.2431 Google Scholar CrossRef Search ADS Elkins, J. S., Johnston, S. C., Ziv, E., Kado, D., Cauley, J. A., & Yaffe, K. ( 2007). Methylenetetrahydrofolate reductase C677T polymorphism and cognitive function in older women. American Journal of Epidemiology , 166, 672– 678. doi: 10.1093/aje/kwm140 Google Scholar CrossRef Search ADS Folstein, M. F., Folstein, S. E., & McHugh, P. R. ( 1975). “Mini-mental state”. A practical method for grading the cognitive state of patients for the clinician. Journal of Psychiatric Research , 12, 189– 198. doi:10.1016/0022-3956(75)90026-6 Google Scholar CrossRef Search ADS Hayden, K. M., Lutz, M. W., Kuchibhatla, M., Germain, C., & Plassman, B. L. ( 2015). Effect of APOE and CD33 on cognitive decline. PLoS One , 10, e0130419. doi: 10.1371/journal.pone.0130419 Google Scholar CrossRef Search ADS Jefferson, A. L., Hohman, T. J., Liu, D., Haj-Hassan, S., Gifford, K. A., Benson, E. M., … Ruberg, F. L. ( 2015). Adverse vascular risk is related to cognitive decline in older adults. Journal of Alzheimer’s Disease , 44, 1361– 1373. doi: 10.3233/JAD-141812 Kaffashian, S., Dugravot, A., Elbaz, A., Shipley, M. J., Sabia, S., Kivimäki, M., & Singh-Manoux, A. ( 2013). Predicting cognitive decline: A dementia risk score vs. the Framingham vascular risk scores. Neurology , 80, 1300– 1306. doi: 10.1212/WNL.0b013e31828ab370 Google Scholar CrossRef Search ADS Kaup, A. R., Nettiksimmons, J., Harris, T. B., Sink, K. M., Satterfield, S., Metti, A. L., … Body Composition, S. ( 2015). Cognitive resilience to apolipoprotein E epsilon4: Contributing factors in black and white older adults. JAMA Neurology , 72, 340– 348. doi: 10.1001/jamaneurol.2014.3978 Google Scholar CrossRef Search ADS Kivipelto, M., Ngandu, T., Laatikainen, T., Winblad, B., Soininen, H., & Tuomilehto, J. ( 2006). Risk score for the prediction of dementia risk in 20 years among middle aged people: A longitudinal, population-based study. The Lancet Neurology , 5, 735– 741. doi: 10.1016/S1474-4422(06)70537-3 Google Scholar CrossRef Search ADS Kleiman, T., Zdanys, K., Black, B., Rightmer, T., Grey, M., Garman, K., … van Dyck, C. ( 2006). Apolipoprotein E epsilon4 allele is unrelated to cognitive or functional decline in Alzheimer’s disease: Retrospective and prospective analysis. Dementia and Geriatric Cognitive Disorders , 22, 73– 82. doi: 10.1159/000093316 Google Scholar CrossRef Search ADS Knopman, D., Boland, L. L., Mosley, T., Howard, G., Liao, D., Szklo, M., … Atherosclerosis Risk in Communities Study, I. ( 2001). Cardiovascular risk factors and cognitive decline in middle-aged adults. Neurology , 56, 42– 48. doi: 10.1212/wnl.56.1.42 Google Scholar CrossRef Search ADS Krabbe, K. S., Mortensen, E. L., Avlund, K., Pilegaard, H., Christiansen, L., Pedersen, A. N., … Bruunsgaard, H. ( 2009). Genetic priming of a proinflammatory profile predicts low IQ in octogenarians. Neurobiology of Aging , 30, 769– 781. doi: 10.1016/j.neurobiolaging.2007.08.013 Google Scholar CrossRef Search ADS Lambert, J. C., Ibrahim-Verbaas, C. A., Harold, D., Naj, A. C., Sims, R., Bellenguez, C., … Amouyel, P. ( 2013). Meta-analysis of 74,046 individuals identifies 11 new susceptibility loci for Alzheimer’s disease. Nature Genetics , 45, 1452– 1458. doi: 10.1038/ng.2802 Google Scholar CrossRef Search ADS Lang, I. A., Llewellyn, D. J., Langa, K. M., Wallace, R. B., Huppert, F. A., & Melzer, D. ( 2008). Neighborhood deprivation, individual socioeconomic status, and cognitive function in older people: Analyses from the English longitudinal study of ageing. Journal of the American Geriatrics Society , 56, 191– 198. doi: 10.1111/j.1532-5415.2007.01557.x Google Scholar CrossRef Search ADS Liaw, A., & Wiener, M. ( 2002). Classification and regression by randomforest. Rnews , 2, 18– 22. Marioni, R. E., Strachan, M. W., Reynolds, R. M., Lowe, G. D., Mitchell, R. J., Fowkes, F. G., … Price, J. F. ( 2010). Association between raised inflammatory markers and cognitive decline in elderly people with type 2 diabetes: The edinburgh type 2 diabetes study. Diabetes , 59, 710– 713. doi: 10.2337/db09-1163 Google Scholar CrossRef Search ADS McCarrey, A. C., An, Y., Kitner-Triolo, M. H., Ferrucci, L., & Resnick, S. M. ( 2016). Sex differences in cognitive trajectories in clinically normal older adults. Psychology and Aging , 31, 166– 175. doi: 10.1037/pag0000070 Google Scholar CrossRef Search ADS McDermott, M., Lisabeth, L. D., Baek, J., Adelman, E. E., Garcia, N. M., Case, E., … Zahuranec, D. B. ( 2017a). Sex disparity in stroke quality of care in a community-based study. Journal of Stroke and Cerebrovascular Diseases , 26, 1781– 1786. doi: 10.1016/j.jstrokecerebrovasdis.2017.04.006 Google Scholar CrossRef Search ADS McDermott, K. L., McFall, G. P., Andrews, S. J., Anstey, K. J., & Dixon, R. A. ( 2017b). Memory resilience to alzheimer’s genetic risk: Sex effects in predictor profiles. The Journals of Gerontology, Series B: Psychological Sciences and Social Sciences , 72, 937– 946. doi: 10.1093/geronb/gbw161 Miles, J. N., Weden, M. M., Lavery, D., Escarce, J. J., Cagney, K. A., & Shih, R. A. ( 2016). Constructing a time-invariant measure of the socio-economic status of U.S. census tracts. Journal of Urban Health: Bulletin of the New York Academy of Medicine , 93, 213– 232. doi: 10.1007/s11524-015-9959-y Google Scholar CrossRef Search ADS Mooijaart, S. P., Sattar, N., Trompet, S., Polisecki, E., de Craen, A. J., Schaefer, E. J., … Westendorp, R. G.; PROSPER Study Group. ( 2011). C-reactive protein and genetic variants and cognitive decline in old age: The PROSPER study. PLoS One , 6, e23890. doi: 10.1371/journal.pone.0023890 Google Scholar CrossRef Search ADS Norton, S., Matthews, F. E., Barnes, D. E., Yaffe, K., & Brayne, C. ( 2014). Potential for primary prevention of Alzheimer’s disease: An analysis of population-based data. The Lancet Neurology , 13, 788– 794. doi: 10.1016/S1474-4422(14)70136-X Google Scholar CrossRef Search ADS Payton, A., Gibbons, L., Davidson, Y., Ollier, W., Rabbitt, P., Worthington, J., … Horan, M. ( 2005). Influence of serotonin transporter gene polymorphisms on cognitive decline and cognitive abilities in a nondemented elderly population. Molecular Psychiatry , 10, 1133– 1139. doi: 10.1038/sj.mp.4001733 Google Scholar CrossRef Search ADS Proust-Lima, C., Amieva, H., Dartigues, J. F., & Jacqmin-Gadda, H. ( 2007). Sensitivity of four psychometric tests to measure cognitive changes in brain aging-population-based studies. American Journal of Epidemiology , 165, 344– 350. doi: 10.1093/aje/kwk017 Google Scholar CrossRef Search ADS RAND. ( 2010). Neighborhood SES Index Data Core User’s Documentation Series 2010, 15. (2010). Retrieved from: http://www.rand.org/health/centers/pophealth/data.html Ryan, C. L., & Bauman, K. ( 2016). Educational Attainment in the United States: 2015 . Sanchez, M. M., Das, D., Taylor, J. L., Noda, A., Yesavage, J. A., & Salehi, A. ( 2011). BDNF polymorphism predicts the rate of decline in skilled task performance and hippocampal volume in healthy individuals. Translational Psychiatry , 1, e51. doi: 10.1038/tp.2011.47 Google Scholar CrossRef Search ADS Schmand, B., Smit, J., Lindeboom, J., Smits, C., Hooijer, C., Jonker, C., & Deelman, B. ( 1997). Low education is a genuine risk factor for accelerated memory decline and dementia. Journal of Clinical Epidemiology , 50, 1025– 1033. doi: 10.1016/S0895-4356(97)00121-2 Google Scholar CrossRef Search ADS Schneider, A. L., Sharrett, A. R., Patel, M. D., Alonso, A., Coresh, J., Mosley, T., … Gottesman, R. F. ( 2012). Education and cognitive change over 15 years: The atherosclerosis risk in communities study. Journal of the American Geriatrics Society , 60, 1847– 1853. doi: 10.1111/j.1532-5415.2012.04164.x Google Scholar CrossRef Search ADS Sherva, R., Tripodis, Y., Bennett, D. A., Chibnik, L. B., Crane, P. K., de Jager, P. L., … Green, R. C.; GENAROAD Consortium; Alzheimer’s Disease Neuroimaging Initiative; Alzheimer’s Disease Genetics Consortium. ( 2014). Genome-wide association study of the rate of cognitive decline in Alzheimer’s disease. Alzheimer’s & Dementia , 10, 45– 52. doi: 10.1016/j.jalz.2013.01.008 Google Scholar CrossRef Search ADS Shih, R. A., Ghosh-Dastidar, B., Margolis, K. L., Slaughter, M. E., Jewell, A., Bird, C. E., … Espeland, M. A. ( 2011). Neighborhood socioeconomic status and cognitive function in women. American Journal of Public Health , 101, 1721– 1728. doi: 10.2105/AJPH.2011.300169 Google Scholar CrossRef Search ADS Shpanskaya, K. S., Choudhury, K. R., Hostage, C. Jr, Murphy, K. R., Petrella, J. R., & Doraiswamy, P. M.; Alzheimer’s Disease Neuroimaging Initiative. ( 2014). Educational attainment and hippocampal atrophy in the Alzheimer’s disease neuroimaging initiative cohort. Journal of Neuroradiology , 41, 350– 357. doi: 10.1016/j.neurad.2013.11.004 Google Scholar CrossRef Search ADS Stern, Y. ( 2012). Cognitive reserve in ageing and Alzheimer’s disease. The Lancet. Neurology , 11, 1006– 1012. doi: 10.1016/S1474-4422(12)70191-6 Google Scholar CrossRef Search ADS Stern, Y., Gurland, B., Tatemichi, T. K., Tang, M. X., Wilder, D., & Mayeux, R. ( 1994). Influence of education and occupation on the incidence of Alzheimer’s disease. JAMA , 271, 1004– 1010. doi:10.1001/jama.1994.03510370056032 Google Scholar CrossRef Search ADS Strobl, C., Boulesteix, A. L., Kneib, T., Augustin, T., & Zeileis, A. ( 2008). Conditional variable importance for random forests. BMC Bioinformatics , 9, 307. doi: 10.1186/1471-2105-9-307 Google Scholar CrossRef Search ADS Strobl, C., Malley, J., & Tutz, G. ( 2009). An introduction to recursive partitioning: Rationale, application, and characteristics of classification and regression trees, bagging, and random forests. Psychological Methods , 14, 323– 348. doi: 10.1037/a0016973 Google Scholar CrossRef Search ADS Tsai, S. J., Hong, C. J., Liu, M. E., Hou, S. J., Yen, F. C., Hsieh, C. H., & Liou, Y. J. ( 2010). Interleukin-1 beta (C-511T) genetic polymorphism is associated with cognitive performance in elderly males without dementia. Neurobiology of Aging , 31, 1950– 1955. doi: 10.1016/j.neurobiolaging.2008.10.002 Google Scholar CrossRef Search ADS Vemuri, P., Lesnick, T. G., Przybelski, S. A., Machulda, M., Knopman, D. S., Mielke, M. M., … Jack, C. R. Jr. ( 2014). Association of lifetime intellectual enrichment with cognitive decline in the older population. JAMA Neurology , 71, 1017– 1024. doi: 10.1001/jamaneurol.2014.963 Google Scholar CrossRef Search ADS Vermeer, S. E., Prins, N. D., den Heijer, T., Hofman, A., Koudstaal, P. J., & Breteler, M. M. ( 2003). Silent brain infarcts and the risk of dementia and cognitive decline. The New England Journal of Medicine , 348, 1215– 1222. doi: 10.1056/NEJMoa022066 Google Scholar CrossRef Search ADS Welsh, K. A., Breitner, J. C. S., & Habib-Magruder, K. M. ( 1993). Detection of dementia in the elderly using telephone screening of cognitive status. Neuropsychiatry Neuropsychology and Behavioral Neurology , 6, 103– 110. Wilson, R. S., Hebert, L. E., Scherr, P. A., Barnes, L. L., Mendes de Leon, C. F., & Evans, D. A. ( 2009). Educational attainment and cognitive decline in old age. Neurology , 72, 460– 465. doi: 10.1212/01.wnl.0000341782.71418.6c Google Scholar CrossRef Search ADS Yaffe, K., Kanaya, A. M., Lindquist, K., Hsueh, W. C., Cummings, S. R., Beamer, B., … Harris, T.; Health ABC Study. ( 2008). PPAR-gamma Pro12Ala genotype and risk of cognitive decline in elders. Neurobiology of Aging , 29, 78– 83. doi: 10.1016/j.neurobiolaging.2006.09.010 Google Scholar CrossRef Search ADS Zahodne, L. B., Glymour, M. M., Sparks, C., Bontempo, D., Dixon, R. A., MacDonald, S. W., & Manly, J. J. ( 2011). Education does not slow cognitive decline with aging: 12-year evidence from the victoria longitudinal study. Journal of the International Neuropsychological Society , 17, 1039– 1046. doi: 10.1017/S1355617711001044 Google Scholar CrossRef Search ADS Zeki Al Hazzouri, A., Haan, M. N., Osypuk, T., Abdou, C., Hinton, L., & Aiello, A. E. ( 2011). Neighborhood socioeconomic context and cognitive decline among older Mexican Americans: Results from the sacramento area latino study on aging. American Journal of Epidemiology , 174, 423– 431. doi:10.1093/aje/kwr095 Google Scholar CrossRef Search ADS © The Author(s) 2018. Published by Oxford University Press on behalf of The Gerontological Society of America. All rights reserved. For permissions, please e-mail: firstname.lastname@example.org. This article is published and distributed under the terms of the Oxford University Press, Standard Journals Publication Model (https://academic.oup.com/journals/pages/about_us/legal/notices)
The Journals of Gerontology Series B: Psychological Sciences and Social Sciences – Oxford University Press
Published: Apr 27, 2018
It’s your single place to instantly
discover and read the research
that matters to you.
Enjoy affordable access to
over 18 million articles from more than
15,000 peer-reviewed journals.
All for just $49/month
Query the DeepDyve database, plus search all of PubMed and Google Scholar seamlessly
Save any article or search result from DeepDyve, PubMed, and Google Scholar... all in one place.
Get unlimited, online access to over 18 million full-text articles from more than 15,000 scientific journals.
Read from thousands of the leading scholarly journals from SpringerNature, Elsevier, Wiley-Blackwell, Oxford University Press and more.
All the latest content is available, no embargo periods.
“Hi guys, I cannot tell you how much I love this resource. Incredible. I really believe you've hit the nail on the head with this site in regards to solving the research-purchase issue.”Daniel C.
“Whoa! It’s like Spotify but for academic articles.”@Phil_Robichaud
“I must say, @deepdyve is a fabulous solution to the independent researcher's problem of #access to #information.”@deepthiw
“My last article couldn't be possible without the platform @deepdyve that makes journal papers cheaper.”@JoseServera