Data-Driven Approaches to Executive Function Performance and Structure in Aging: Integrating Person-Centered Analyses and Machine Learning Risk Prediction

Abstract

Executive function (EF) performance and structure in non-demented aging are frequently examined with variable-centered approaches. Person-centered analytics can contribute unique information about classes of persons by simultaneously considering EF performance and structure. The risk predictors of these classes can then be determined by machine learning technology.

Objective:

Using data from the Victoria Longitudinal Study we examined two goals: (1) detect different underlying subgroups (or classes) of EF performance and structure and (2) test multiple risk predictors for best discrimination of these detected subgroups.

Method:

We used a classification sample (n = 778; M age = 71.42) for the first goal and a prediction sub-sample (n = 570; M age = 70.10) for the second goal. Eight neuropsychological measures represented three EF dimensions (inhibition, updating, shifting). Fifteen predictors represented five domains (genetic, functional, lifestyle, mobility, demographic).

Results:

First, we observed two distinct classes: (a) lower EF performance and unidimensional structure (Class 1) and (b) higher EF performance and multidimensional structure (Class 2). Second, Class 2 was predicted by younger age, more novel cognitive activity, more education, lower body mass index, lower pulse pressure, female sex, faster balance, and more physical activity.

Conclusions:

Data-driven modeling approaches tested the possibility of an EF aging class that displayed both preserved EF performance levels and sustained multidimensional structure. The two observed classes differed in both performance level (lower, higher) and structure (unidimensional, multidimensional). Machine learning prediction analyses showed that the higher performing and multidimensional class was associated with multiple brain health-related protective factors.

The dynamics of human cognition and action are monitored by a set of mental control processes referred to as executive functions (EFs; Miyake & Friedman, 2012). The most common approach to EF and aging research may be characterized as variable-centered (i.e., examining relations among variables across people; Laursen et al., 2006). However, a complementary approach, known as person-centered (i.e., identifying classes of persons based on their similarities across variables) may provide novel information about EF performance (level) and structure (unidimensional or multidimensional factor) in aging. For research on EF performance and change, the main goals have been to establish the extent to which (a) EF performance declines in older adults as compared to younger adults and (b) there are variable patterns in the rate of decline among non-demented aging individuals (Allain et al., 2005; Hull et al., 2008; Lin et al., 2017; Sha et al., 2019). However, most of our present knowledge about the structure of EF has been produced by a variety of variable-centered approaches, including confirmatory and exploratory factor analysis (Bock et al., 2019; Kievit et al., 2014; McFall et al., 2014; Vaughan & Giovanello, 2010). To our knowledge, person-center approaches have not been used to study EF structure in aging. Furthermore, EF performance and structure are frequently examined separately. We adopt a data-driven person-centered approach such that classes of persons who share similar patterns (e.g., performance and structure) can be identified. The integration of this approach to the study of EF performance and structure may provide (a) more complete understanding of the processes and patterns of EF and (b) direct analyses and comparison of EF performance and structure.

How might person-centered analytic approaches provide additional, or even unique, information about performance and structural changes in EF aging? Using a variable-centered approach, research has shown that, from early childhood, EFs exhibit a unidimensional (one factor) structure that then evolves into a multidimensional (two or three factors) structure in older children, adolescents, and young adults (Chevalier & Clark, 2017; Lee et al., 2013; Miyake et al., 2000; Wiebe & Karbach, 2017; Wiebe et al., 2011). Extending this research into adulthood and aging, EFs appear to undergo a change in structure that is represented again by a unidimensional factor in both non-demented and impaired aging (Adrover-Roig et al., 2012; de Frias et al., 2006, 2009; Li et al., 2017; McFall et al., 2014; McFall et al., 2017). The unity/diversity model (Friedman & Miyake, 2017) indicates that cognitively normal individuals are equipped with the same EF processes, but there is enough variability to produce important individual differences in test performance or structural characteristics. An implication for non-impaired cognitive aging is that many individuals may show an expected unity of EFs (e.g., unidimensional structure) whereas other individuals may retain a diversity of EFs (e.g., multidimensional structure). Some preliminary evidence of the latter has been reported with variable-centered approaches (e.g., de Frias et al., 2009; Hedden & Yoon, 2006; Hull et al., 2008; Kievit et al., 2014). It is possible that person-centered approaches may provide more conclusive evidence about individual patterns and predictors of dimensionality of EF structure in aging. For example, such approaches may lead to insights about the extent to which differential aging changes in EF performance are related to aging changes in EF structure.

In the course of recent EF research, multiple risk factor predictors of cross-sectional EF deficits, declining EF performance, or emerging EF impairment have been identified. Many of these factors are potentially modifiable, including education (Chen et al., 2019; Dorbath et al., 2013), physical and cognitive activity (Blasko et al., 2014; Thibeau et al., 2016), pulse pressure (McFall et al., 2014; Raz et al., 2011), and body mass index (Anstey et al., 2011; Bohn et al., 2020). Non-modifiable genetic factors involve polymorphisms such as the insulin degrading enzyme (McFall et al., 2013; Zhang et al., 2013), catechol-O-methyltransferase (Perkovic et al., 2018; Holtzer et al., 2010), brain-derived neurotrophic factor (Sapkota et al., 2015; Toh et al., 2018), and apolipoprotein E (Sapkota et al., 2017; Wisdom et al., 2011). Non-modifiable demographic factors involve age (Li et al., 2017; McCarrey et al., 2016) and sex (male or female; Mansouri et al., 2016; Zaninotto et al., 2018). Other risk factors that have been related to differential EF aging decline include grip strength (Boyle et al., 2009; Sternäng et al., 2015), peak expiratory flow (Qiao et al., 2020; Vidal et al., 2013), gait (Mirelman et al., 2012; Thibeau et al., 2019), and balance (Amboni et al., 2013; Kearney et al., 2013). These risk indicators present reasonable targets for predictions of differences in both EF performance and structure. Most of the above-cited research is conducted at the candidate risk factor level and thus focuses on single or very few predictor variables. New quantitative technologies (e.g., machine learning algorithms) permit simultaneous testing of multiple potential predictors in a quantitatively competitive context. These approaches selectively discriminate predictors of relatively more potential importance from those of relatively less importance, given internal validation procedures (Caballero et al., 2021; Lovatti et al., 2019; McFall et al., 2019).

Although considerable research has been conducted on selected independent predictors of EF performance and change in aging, we are unaware of research that coordinates the two forms of data-driven analytics. First, we use factor mixture modeling (FMM) as a person-centered approach to test for performance and structural differences in a sample of non-demented aging individuals. The novel outcome would be to establish classes of persons based on simultaneous consideration of both EF performance and structure. The conceptual significance would be to denote if there is a direct or indirect relationship between EF performance and structure. Second, we use random forest analysis (RFA), a machine learning prediction model with the out-of-bag error embedded feature for validation, to determine the most important predictors (in this dataset) of expected classes as observed in the first analysis. Specifically, we assemble and compare 15 predictors in a competitive computational context to determine their relative importance in discriminating classes with a specific performance and structural pattern.

Accordingly, we pursued two research goals. For the first goal, we used FMM to produce classes in a large cross-sectional sample of normal aging adults. We expected to find distinct classes of EF performance and that the observed classes would be represented by either a unidimensional or multidimensional structure. For the second goal, we used RFA to test and compare 15 risk factor predictors, independently related to aspects of EF performance and change, to identify those that objectively discriminate classes produced by the first analysis.

Method

Participants

The participants were community-dwelling adults from the Victoria Longitudinal Study (VLS), an ongoing large-scale, multi-sample, longitudinal study that examines multiple aspects of human aging including genetic, biomedical, functional, neuropsychological, and lifestyle (Dixon & de Frias, 2004). Originally, participants were recruited through the public media and requests from community groups. Participants were paid nominal fees for their participation and provided written informed consent. Data collection procedures were in full and certified compliance with the University of Alberta Human Research Ethics Board. Using standard procedures (i.e., McFall et al., 2014), we assembled cross-sectional data from the 2000–2016 period of the VLS archives. The data included adults initially aged 55–95 years. As displayed in the consort flow diagram (Figure 1), we applied systematic exclusion rules to select both classification (for research goal 1) and prediction (for research goal 2) samples from a larger VLS source sample.

Figure 1. Consort Diagram.

Note. Participants from the VLS underwent multiple exclusionary criteria. The classification sample served the first goal. The prediction sample served the second goal.

The source sample included 914 persons (M age = 71.91, SD = 9.18, range = 53.24 – 100.16, 66.2% female, M years of education = 15.09). Participants were excluded if they had (a) full EF data missing (n = 23), (b) reported diagnosis of mild to very serious Alzheimer’s disease or other forms of cognitive impairment and dementia (n = 6), (c) self-reported history of very serious epilepsy, depression, and head injury (e.g., in a coma, hospitalized; n = 30), (d) self-reported moderately serious to very serious stroke (any region; n = 29), (e) self-reported history of moderate-to-severe Parkinson’s disease (n = 7), (f) Mini-Mental Status Examination (MMSE) score less than 24 (n = 35), and (g) use of anti-psychotic medication (n = 6). We thereby established the classification sample consisting of 778 participants (M age = 71.42, SD = 9.07, range = 53.24 – 95.25, 66.5% female, M years of education = 15.20). The race/ethnicity information is as follows: 98.8% Caucasian; 0.9% Pacific Islander; 0.2% Asian; 0.2% Metis. Table 1 shows descriptive statistics for this sample, including mean performance for each EF measure used.

Table 1.

Descriptive Statistics for Classification Sample

N	778
Age (years)	71.42 (9.07); 53.24 – 95.25
Sex (% female)	66.6
Education (years)	15.20 (3.00); 5.00 – 27.00
Hayling a	5.43 (1.50); 1.00 – 10.00
Stroop (interference index) b	1.28 (.75); −.73 – 6.14
Color Trails (seconds) b	98.01 (33.87); 46.67 – 266.22
Brixton a	4.72 (2.18); 1.00 – 10.00
Letter Series a	11.33 (4.44); 0 – 20.00
Letter Sets a	8.36 (2.89); 0 – 14.00
Computational Span a	3.05 (1.27); 0 – 7.00
Reading Span a	2.88 (1.02); 0 – 6.00

Note. Results presented as Mean (Standard Deviation) and Range unless otherwise stated. The maximum score for the measures with a defined upper limit is as follows: Hayling 10, Brixton 10, Letter Series 20, Letter Sets 15, Computational Span 7, Reading Span 7.

^aScaled score.

^bLower scores indicate better performance.

Beginning with the classification sample, we then established a prediction sample. A subset of the classification sample contributed genetic data during a later phase (2009–2011). As genetic risk was an important domain of prediction in this study, we established a prediction sample comprised of a complete subset of genotyped participants (n = 570; M age = 70.10, SD = 8.50, range = 53.24 – 95.25, 66.5% female; M years of education = 15.32). Table 2 shows descriptive statistics for the prediction sample, including the risk factor predictor variables.

Table 2.

Descriptive Statistics for Prediction Sample

N	570
Age (years)	70.10 (8.50); 53.24 – 95.25
Sex (% female)	66.5
Education (years)	15.32 (2.96); 5.00 – 24.00
BDNF (% Met+)	34.9
IDE (% G−)	13.3
APOE (% ε4+)	23.0
COMT (% Val+)	77.4
Pulse Pressure (mmHg)	52.06 (10.30); 32.13 – 99.25
Body Mass Index (kg/m2)	26.91 (4.17); 14.95 – 48.61
Gait (seconds)	6.33 (1.64); 3.38 – 21.81
Balance (seconds)	2.78 (1.01); .78 – 12.19
Peak Expiratory Flow (litres/minute)	426.02 (119.06); 0 – 770.00
Grip Strength (kg/force)	29.63 (9.44); 9.50 – 59.25
Everyday Physical Activity a	15.95 (5.15); 0 – 31.00
Everyday Novel Cognitive Activity a	75.81 (16.78); 20.00 – 130.00

Note. Results presented as Mean (Standard Deviation) and Range unless otherwise stated. The maximum score for the measures with a defined upper limit is as follows: Peak Expiratory Flow 800, Grip Strength 100, Everyday Physical Activity 32, Everyday Novel Cognitive Activity 216. BDNF = Brain Derived Neurotrophic Factor; IDE = Insulin Degrading Enzyme; APOE = Apolipoprotein E; COMT= Catechol-O-Methyl Transferase; Met+ = risk allele combinations; G−= risk allele combination; ε4+ = risk allele combinations. Val+ = risk allele combinations; mmHg = millimetre of mercury; kg/m² = kilograms/metres²; kg/force = kilograms/force.

^aScaled score.

Executive Function Measures

For the first goal, we used eight standard neuropsychological measures for the three EF dimensions: two each for inhibition (Hayling, Stroop) and updating (Computational Span, Reading Span) and four for shifting (Brixton, Color Trails, Letter Series, Letter Sets). Information about these tests and their psychometric properties are reported in multiple studies (Bielak et al., 2006; de Frias & Dixon, 2014; de Frias et al., 2009; McFall et al., 2014; McFall et al., 2013; Sapkota et al., 2015; Sapkota et al., 2017). More information is provided in Supplemental Material.

DNA Extraction and Genotyping

Saliva samples were collected according to Oragene DNA Genotek technology protocol. Genotyping was carried out by using a Polymerase Chain Reaction Restriction Fragment Length Polymorphism strategy to analyze the allele status of insulin degrading enzyme (IDE; rs6583817), catechol-O-methyltransferase (COMT; rs4680), brain-derived neurotrophic factor (BDNF; rs6265), and apolipoprotein E (APOE; determined by the combination of the SNPs rs429358 and rs7412). A detailed description of DNA extraction and genotyping is provided elsewhere (McFall et al., 2013; Sapkota et al., 2017; Thibeau et al., 2016).

Fifteen Risk Factor Predictors

For the second goal, a total of 15 predictors from five domains were used to discriminate classes of persons based on EF performance and structure. The pool of predictors was based on reviews of prominent risk factors for general cognitive aging decline and impairment (Dixon & Lachman, 2019; Livingston et al., 2017) and specifically for EF decline and unidimensional structure, in the context of the VLS data base.

Genetic

We included the following genetic polymorphisms from DNA extraction and genotyping: (a) IDE, (b) COMT, (c) BDNF, and (d) APOE. We conducted a dichotomous genotype categorization based on the presence or absence of the genotype risk allele. For IDE genotype, we used G+ (non-risk; composed of the GG and GA allele combinations) and G- (risk; composed of the AA allele combination). For COMT genotype, we used Val- (non-risk; composed of the Met/Met allele combination) and Val+ (risk; composed of the Val/Val and Val/Met allele combinations). For BDNF genotype, we used Met- (non-risk; composed of the Val/Val allele combination) and Met+ (risk; composed of the Met/Met and Val/Met allele combinations). For APOE genotype, we used ε4− (non-risk; composed of ε2ε2, ε2ε3, ε3ε3 allele combinations) and ε4+ (risk; composed of ε4ε4 and ε3ε4 allele combinations). The genotype frequencies for 3 of the genotypes did not differ significantly from Hardy-Weinberg (HW) equilibrium: COMT [χ² = 2.93 (1), p > .05]; BDNF [χ² = 1.47 (1), p > .05]; APOE [χ² = 0.71 (1), p > .05]. We found a significant departure from HW equilibrium for IDE: [χ² = 62.59 (1), p < .05]. As we have used this polymorphism in previous research (Thibeau et al., 2016), we included it for replication purposes in the present analyses.

Functional

Functional markers included (a) pulse pressure [PP; equals systolic blood pressure (BP) - diastolic BP, in mmHg] based on an average of eight BP readings, (b) body mass index (BMI; equals weight/height ², in kilograms/meters²), (c) peak expiratory flow (PEF; largest volume of air expired over three attempts, in litres/minute), and (d) grip strength (average hand strength, in kilograms/force).

Lifestyle

Lifestyle factors were (a) everyday physical activity (based on 4 self-report questions), and (b) everyday novel cognitive activity (27 self-report questions). These lifestyle variables are part of the VLS Activities Lifestyle Questionnaire (see Runge et al., 2014) and are based on a 9-point scale (never = 0, daily = 8) that rates the frequency of participation.

Mobility

Mobility markers (Thibeau et al., 2019) were (a) balance, measured as timed turn (360-degree turn, in seconds) and (b) gait, measured as timed walk (20 feet, in seconds).

Demographic

Demographic factors included (a) education (total years), (b) age (in years), and (c) sex (male or female).

Statistical Analyses

We used Mplus 8.2 (Muthén & Muthén, 2017) and data from the classification sample for the first goal. We used R 3.3.2 (R Development Core Team, 2015) and data from the prediction sample for the second goal. Less than 2% of data were missing and estimated in the classification sample. We used full information maximum likelihood (FIML) with robust estimator, an approach that is less dependent on the multivariate normality assumption and produces parameter estimates correcting for standard errors and chi-square test statistics (Li, 2016). Missing values in predictors from the prediction sample were handled with the missForest package in R. The missForest algorithm uses a nonparametric imputation method, fits a random forest to the observed data, and then predicts the missing data for each variable (Stekhoven & Bühlmann, 2011). For our prediction sample, the following characteristics for missing data were observed: M percentage = 1.2% (range = 0.4% – 4.0%). The out-of-bag error was as follows: normalized root mean squared error = 0.23, proportion of falsely classified entries = 0.08.

First Goal: Classification of EF Performance and Structure

Factor Mixture Modeling (FMM) was used to classify persons based on both EF performance and structure. FMM is a hybrid of latent class (or latent profile for continuous data) and factor analysis (Lubke & Muthén, 2007). Within the FMM model, the latent class variable assigns each person into classes and the latent continuous factor models the heterogeneity of the construct (i.e., EF) within the latent class (Clark et al., 2013; Masyn et al., 2010). The factor loadings, factor means, factor (co)variance matrix, intercepts, and residual variances have the potential to be class-specific. Therefore, the factor structure of the model can be different in each class. Consequently, FMM is decomposed in several parts. Part 1 is the common factor model and is expressed as:

$${\mathbf{y}}_{\text{i}}=\mathbf{v}+{\mathit{\Lambda }}_y{\mathit{\eta }}_i+{\mathbf{T}}_{\text{y}}{\mathbf{X}}_{\text{i}}+{\mathit{\epsilon }}_i,$$

$${\mathit{\eta }}_i={\mathbf{T}}_n{\text{X}}_{\mathrm{i}}+{\mathit{\zeta }}_i$$

In the first equation, y_i is the participant’s i score on the observed variable y; ѵ represents the regression intercepts; Λ_y is the factor loadings; scores on the underlying factors are denoted η; Ƭ_y is the regression weights; covariates are denoted as X; ε is the regression residuals. For the second equation, the factor scores η are regressed on covariates X with regression weights Ƭ_n. Furthermore, ζ_i is the residual factor score (i.e., part of the factor scores not explained by the covariates; see Lubke & Muthén, 2005).

In part 2 of the FMM, the above equations are extended by regressing the factor scores on the latent class variable C. Including this extension gives the following:

$${\mathbf{y}}_{\text{ik}}={\mathbf{v}}_{\text{k}}+{\mathit{\Lambda }}_{yk}{\mathit{\eta }}_{ik}+{\mathbf{T}}_{\text{yk}}{\mathbf{X}}_{\text{i}}+{\mathit{\epsilon }}_{ik},$$

$${\mathit{\eta }}_{ik}=\mathbf{A}{\mathbf{C}}_{\text{i}}+{\mathbf{T}}_{nk}{\text{X}}_{\text{i}}+{\mathit{\zeta }}_{ik}$$

In these equations, the k subscript is attached to the parameters that may differ across classes (k) and to the random variables with class specific distributions (e.g., means, covariance). A represents the intercepts of the factors for each class (Lubke & Muthén, 2005).

Our approach involved systematically testing and selecting the best class models (latent profile analysis), best factor structure (factor analysis), and best FMM model. As in previous work testing latent classes, we selected models with low Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), −2 log likelihood (−2LL), and classes with substantial (>10%) prevalence (Caballero et al., 2021; McFall et al., 2019; McDermott et al., 2017). We used high entropy value (>.70) to indicate model classification accuracy. The 10% criterion was expected to yield at least 100 participants to accommodate further prediction analyses. The best factor models were determined by (a) comparative fit index (CFI; value of ≥ .95 is good and ≥ .90 is adequate fit), (b) root mean square error of approximation (RMSEA; value of ≤ .05 is good and ≤ .08 is adequate fit), and (c) standardized root mean square residual (SRMR; good fit determined by a value of ≤ .08) (Kline, 2011; Little, 2013).

After selecting the number of classes and factors, a total of 5 sequentially nested (less restrictive) FMM models were fitted to test measurement invariance of EF across classes (see Table 3). Model selection was based on fit indices, including: AIC, Sample-Size Adjusted BIC (SSABIC), Lo-Mendell-Rubin test (LMR), and Bootstrapped Likelihood Ratio Test (BLRT; see Bernstein et al., 2013; Clark et al., 2013). A good fit for these tests produces significance, p < .05, indicating better fit of models with k class than with k-1 class. The analyses allowed us to observe class-specific factor patterns of EF structure.

Table 3.

Factor Mixture Model (FMM) Specifications

Model	Description
FMM-1	Class-varying factor means (i.e., different EF performance); class-invariant factor loadings, intercepts, and residuals (i.e., the same EF structure); factor (co) variance fixed to zero (i.e., no within-class variability).
FMM-2	Class-varying factor means and (co)variance; class-invariant factor loadings, intercepts, and residuals (i.e., strict invariance).
FMM-3	Class-varying factor means and (co) variance; class-invariant factor loadings and intercepts; class-varying residuals (i.e., scalar invariance).
FMM-4	Class-varying factor (co)variance; class-invariant factor loadings; class-varying intercepts and residuals (i.e., metric invariance); latent factor means fixed to zero.
FMM-5	Class-varying factor loadings, intercepts and residuals (i.e., configural invariance); latent factor means and (co) variance fixed to zero.

Note: Each model was systematically specified and evaluated across classes and factors.

Second Goal: Risk Factor Predictors Discriminating Classes of EF Performance and Structure

Random Forest Analysis (RFA) was used to determine important predictors that discriminated classes of persons based on EF performance and structure. RFA is a machine learning analytic technique specifically applicable to biomarker prediction analyses, especially when multiple predictors are examined in a quantitatively competitive context. We have deployed this technology in several recent studies on brain aging and dementia (e.g., Caballero et al., 2021; Drouin et al; 2020; McFall et al., 2019; Sapkota et al., 2018). Specifically, RFA is an ensemble of decision trees constructed from a training dataset that is internally validated to yield a prediction of an outcome (such as differential clinical status) and indicate relative importance of tested predictors (Boulesteix et al., 2012). The computation of RFA involves the creation of an uncorrelated forest of trees, the prediction of which is more accurate than that of any individual tree. Each tree of the random forest is built based on a bootstrap sample drawn randomly from the original dataset using an established splitting criterion known as the Gini impurity (Couronné et al., 2018). At each split, the splitting criterion selects the splitting predictor from a randomly selected subset of predictors; the subset of predictors is different at each split (Boulesteix et al., 2012; Mao & Wang, 2012). As such, RFA provides a computationally competitive environment for evaluating the relative predictive strength of a larger set of tested factors (Lovatti et al., 2019). In this way, RFA can be used to identify leading or important predictors of clinical group differences, concurrent outcomes, or future observations.

RFA was conducted using the “party” package (Hothorn et al., 2006) with out-of-bag error estimation as is appropriate for this type of data (Janitza & Hornung, 2018). A recommended procedure in machine learning and other prediction analyses is to have independent datasets for discovery and validation. However, independent and corresponding datasets are not always possible and a separate dataset was not available for this study. However, RFA includes procedures for internal validation and we performed those procedures in this study. Specifically, the out-of-bag error in RFA is designed for use in internal validation. Briefly, the out-of-bag error is the average error frequency obtained when the observations in the dataset are predicted using the out-of-bag trees (Boulesteix et al., 2012). When the random forest model is being trained using bootstrap aggregation, each new tree is fit from a bootstrap sample of the training observations. The out-of-bag error is calculated using predictions from the trees that do not contain training observations in their respective bootstrap sample. Thus, this procedure provides an established and recommended internal validation dataset for these trees.

We met the following assumptions: (a) uncorrelated trees and (b) no multicollinearity between predictors. Each forest was comprised of a number of ntrees, sufficient for good model stability, and an optimal mtry number of predictors at each potential split. Generally, mtry is set at $\sqrt{\#ofpredictors}$ (Genuer et al., 2010). Permutation accuracy importance was used to assess relative level of importance with the cforest function in the “party” package of R. Specifically, we used the mean decrease in accuracy (MDA) to quantify the importance of each tested predictor by measuring changes in prediction accuracy (Calle & Urrea, 2010) with and without it in the model. This measure is defined as the difference between the out-of-bag error resulting from a dataset acquired through random permutation of the predictor of interest and the out-of-bag error of the original dataset (Boulesteix et al., 2012). Permutation of an important predictor increases the out-of-bag error, which leads to a high importance score. These importance scores determine the magnitude of a variable’s effect and how much predictive power the model loses if the variable is removed from the model (Ishwaran & Lu, 2019). As such, MDA does not require or provide standard statistical indicators such as standard errors, regression coefficients, or confidence intervals to assess precision or accuracy of the prediction estimate. Model strength was assessed as the area under the receiver operation characteristic curve (c-statistic), with values closer to 1 indicating better model strength (see Hajian-Tilaki, 2013). Those variables with negative, zero, or small positive values were not considered important. Descriptive ranking of the predictor variables was used to define importance (McDermott et al., 2017; McFall et al., 2019; Strobl et al., 2009). Direction of predictor effects was determined using correlation analyses.

Results

First Goal: Classification of EF Performance and Structure

Using FMM, we established classes differing in both EF performance and structure. Table 4 shows the three-step FMM procedure with associated results. Results revealed three classes and both a one- and two-factor models to be integrated in the analyses. A maximum of three classes was indicated as four-class models included classes with <10% of the sample and contained non-significant LMR and BLRT p values. Therefore, we tested FMM models based on the following characteristics: (a) two classes, one factor; (b) three classes, one factor; (c) two classes, two factors; (d) three classes, two factors. As can be seen in the table, results showed the FMM-4 model with two classes and two factors to be the best fitting model, AIC = 15681.14, SSABIC = 15743.35, entropy = .80, LMR and BLRT p < .001. FMM-5 models were either unidentified or did not replicate; solutions to models FMM-2 to FMM-4 for three classes and two factors were either not positive definite or were not replicated, suggesting over-extracting information and over-parameterization.

Table 4.

Model Comparison Results

Model	−2LL	Par.	AIC	BIC	SSABIC	ENTROPY	LMR (p-value)	BLRT (p-value)	RMSEA	CFI	SRMR
Latent Profile Analysis
One-Class	−8739.24	16	17510.48	17584.99	17534.18	-	-	-	-	-	-
Two-Class	−8180.53	25	16411.07	16527.48	16448.10	.77	0	0	-	-	-
Three-Class	−8039.24	34	16146.48	16304.81	16196.84	.74	.001	0	-	-	-
Four-Class	−7956.64	43	15999.28	16199.52	16062.97	.78	.34	0	-	-	-
Factor Analysis
One-Factor	−7969.34	26	15990.69	16111.76	16029.19	-	-	-	.02	.99	.02
Two-factor	−7968.55	27	15991.11	16116.84	16031.10	-	-	-	.02	.99	.02
Factor Mixture Analysis
2-Class, 1-Factor
FMM-1	−8180.53	25	16411.07	16527.49	16448.10	.83	.24	0	-	-	-
FMM-2	−8006.45	27	16066.91	16192.64	16106.90	.32	.002	0	-	-	-
FMM-3*	−7878.58	34	15825.16	15983.49	15875.52	.78	.01	0	-	-	-
FMM-4*	−7841.66	41	15765.33	15956.25	15826.06	.80	0	0	-	-	-
3-Class, 1-Factor
FMM-1	−8058.47	27	16170.94	16296.67	16210.93	.75	0	0	-	-	-
FMM-2*	−8004.30	29	16066.60	16201.65	16109.56	.59	.04	.03	-	-	-
FMM-3**	−7802.90	44	15693.80	15898.70	15758.98	.79	.002	0	-	-	-
FMM-4**	−7768.64	58	15653.29	15923.38	15739.11	.79	.17	0	-	-	-
2-Class, 2-Factor
FMM-1	−8180.53	25	16411.07	16527.49	16448.10	.83	0	0	-	-	-
FMM-2	−7954.23	31	15970.45	16114.81	16016.37	.48	.008	0	-	-	-
FMM-3*	−7849.64	36	15771.88	15938.93	15824.61	.79	.006	0	-	-	-
FMM-4* a	−7798.57	42	15681.14	15876.72	15743.37	.80	0	0	-	-	-
3-Class, 2-Factor
FMM-1	−8056.01	28	16168.18	16298.57	16209.65	.75	0	0	-	-	-

Note. FMM-1 = Factor mixture model 1; FMM-2 = Factor mixture model 2; FMM-3 = Factor mixture model 3; FMM-4 = Factor mixture model 4; −2LL = −2 log likelihood; Par. = number of estimated parameters; AIC = Akaike information criterion; BIC = Bayesian information criterion; SSABIC = sample-size adjusted BIC; LMR = Lo-Mendell-Rubin test; BLRT = Bootstrapped Likelihood Ratio Test; RMSEA = Root Mean Square Error of Approximation; CFI = Comparative Fit Index; SRMR = Standardized Root Mean Square Residual.

^aPreferred model.

^*Variance fixed in 1 class at 0 for model identification.

^**Variance fixed in 2 classes at 0 for model identification.

In the identified solution, the two classes had different patterns and characteristics. Specifically, whereas a smaller class (Class 1; n = 151, 19.7%; M age = 79.19, SD = 6.97, range = 58.15 – 95.25) had lower EF performance, a larger class (Class 2; n = 627, 80.3%; M age = 69.55, SD = 8.51, range = 53.24 – 90.32) had higher EF performance. For Class 1, the variances for both factors were negligible and subsequently fixed at 0. Therefore, this class was characterized as reflecting a unidimensional structure. In contrast, Class 2 had significant variance in one factor (shifting, σ² = .09, p < .001) and modest and insignificant variance in the other factor (inhibition + updating, σ² = .01, p = .39). The two factors were correlated (r = −.61, p < .001) in Class 2. Consequently, this class was characterized as reflecting a multidimensional structure. Table 5 provides the standardized factor loadings for each of the EF indicator across the two factors with their standard errors and significance (p) values. Figure 2 shows the distinction of these classes based on their performance on each of the eight EF indicators.

Table 5.

Standardized Factor Loadings (SEs) for EF Indicators Across the Two Factors

Indicator	Factor 1 (Inhibition + Updating)		Factor 2 (Shifting)
	λ(SE)	p	λ(SE)	P
Hayling	.09 (.05)	p = .084	—	—
Stroop	−.26 (.05)	p < .001	—	—
Computational Span	.71 (.04)	p < .001	—	—
Reading Span	.71 (.03)	p < .001	—	—
Brixton	—	—	−.35 (.05)	p < .001
Color Trails	—	—	.48 (.04)	p < .001
Letter Series	—	—	−.81 (.03)	p < .001
Letter Sets	—	—	−.73 (.03)	p < .001

Note: λ = standardized factor loadings.

Figure 2. Executive Function Performance on Each of the Eight Indicators by Class.

Note. The person-centered data-driven analyses produced two observed classes. The figure displays the clear separation across all eight neuropsychological indicator scores for these two classes of older adult participants. The scale on the Y axis represents the range of performance scores. Class 1 (blue) had lower EF performance (all negative indicator means) compared to Class 2 (orange) which had higher EF performance (all positive indicator means). All estimates significant at p < .001, except for Letter Series in Class 2 where p = .010. Class 1 SE range: 0.08 – 0.16. Class 2 SE range: 0.05 – 0.10.

Second Goal: Risk Factor Predictors Discriminating Classes of EF Performance and Structure

We used RFA to compute the relative predictive importance of 15 brain aging-related risk and protective markers in discriminating classes of persons based on EF performance and structure. The model classification performance (c-statistic) was 0.80, 95% CI [.75 – .85], mtry = 4, ntree = 5000. The c-statistic value represented good model performance (Hajian-Tilaki, 2013). Figure 3 shows the predictors in the order of importance, with the first eight at the top and to the right of the vertical line having the best permutation accuracy importance. The eight important predictors represented four of the five domains. Specifically, the important domains and predictors were from the demographic (age, education, sex), lifestyle (everyday novel cognitive activity, everyday physical activity), functional (BMI, PP), and mobility (balance) domains. For directionality discriminating Class 1 from Class 2, the former class was predicted by younger age, more everyday novel cognitive activity, higher education, lower BMI, lower PP, female sex, faster balance, and more everyday physical activity.

Figure 3. Relative Importance of Predictors of Class 1 Versus Class 2.

Note. Predictors to the right of the dashed vertical line are determined to be more important in discriminating the classes than those to the left of the line. On the X axis, Variable Importance refers to the mean decrease in accuracy (negative direction). This metric quantifies the importance of the predictor variable by measuring changes in prediction accuracy. Importance values are presented in scientific notation for the top 8 predictors.

Discussion

Previous perspectives on EF aging have indicated the possibility of a subgroup of aging adults who can be characterized as displaying both relatively preserved levels of EF performance and a sustained multidimensional structure. Our sequence of data-driven modeling approaches tested this possibility directly. Specifically, we examined and integrated three important characteristics in the study of EF in aging: performance, structural dimensionality, and predictors of EF classes. We implemented two research goals: (1) classification of EF performance and structure and (2) risk factor predictors discriminating classes of EF performance and structure. We included a larger classification sample for the first goal and a smaller subset prediction sample for the second goal.

First Goal

We implemented person-centered analytics (FMM) to directly investigate the relationship between EF performance and structure in aging. In the EF and aging literature, performance and structure are frequently examined separately via variable-centered approaches. Currently, there is mixed evidence about the structure of EF in non-demented aging. Whereas some researchers report a unidimensional structure (Adrover-Roig et al., 2012; Salthouse et al., 2003; Thibeau et al., 2016), others report a multidimensional structure (Hedden & Yoon, 2006; Hull et al., 2008; Kievit et al., 2014). With a person-centered approach, we identified classes of persons with similar patterns (performance and structure) interdependently to test the extent to which differential aging EF performance is related to differential aging changes in EF structure.

Within two-factor solutions, we found two classes varying in prevalence (relative size of class) and important class characteristics (performance, both mean and range/variability; see Figure 2). First, in the smaller class (Class 1; n = 151), individuals performed at a relatively low EF level (all negative standardized mean scores and ranges) and results supported the interpretation of a unidimensional structure—a combination expected theoretically. Second, in the larger class (Class 2; n = 627), individuals demonstrated a relatively higher level (all positive standardized mean scores and ranges) of EF performance as well as a sustained multidimensional structure—a combination likely associated with successful EF aging. Although the present study was data-driven—and therefore not designed to test specific hypotheses associated with theories of EF aging—we note that some aspects of the results pertain to an important theoretical perspective, the dedifferentiation approach to EF and aging. Briefly, this approach indicates that with aging, EF structure (and possibly broader cognitive domains) may change from multidimensional (in early adulthood) to unidimensional (less differentiated) in later adulthood (de Frias et al., 2006, 2009; Friedman & Miyake, 2017; Hülür et al., 2015; Miyake et al., 2000; Miyake & Friedman, 2012; Wiebe & Karbach, 2017). Age-related dedifferentiation probably results from biological decay, cultural influences, and lifelong experiences. Notably, dedifferentiation may correspond to the recruitment of larger and overlapping brain areas for a particular task (Bock et al., 2019) and be characterized by declining EF performance. Our results contribute to this perspective as follows. Using new person-centered and data-driven analyses, we observed a novel class of older adults who display (and may have maintained) both higher EF performance and differentiated structure. This class was differentiated objectively from a more typical older adult subgroup who present lower EF performance and who display an agingtypical dedifferentiated structure. The detection of subtle but important factor structure differences accompanied by performance stability can be attributed to the sensitivity of the present FMM approach as applied to complex data (Clark et al., 2013; Masyn et al., 2010).

These results, based on person-centered (data-driven) analyses, provide a complementary perspective to the unity/diversity model, which was developed with variable-centered analyses (Friedman & Miyake, 2017). Importantly, both approaches produce unequivocal evidence for substantial individual differences in EF performance, both in younger and older adults. However, the unity/diversity model also provides a framework for accommodating aging-related results that show EF abilities consolidating into unity (unidimensional) and diversity (multidimensional) patterns, with the notion that these patterns may be associated with differential performance profiles. Our cross-sectional results show support for this notion and indicate that, at least for cognitively normal older adults, performance and structure are not decoupled —and this may be detected by the present data-driven technologies. It is still possible—and not known from the present study—that as participants in the higher performing and multidimensional class age, the expected unity dynamic may emerge with the one-factor solution. This interesting question will require longitudinal data. We note, however, that the present sample, although cross-sectional, included a broad band of older adults (ages 55–95).

Second Goal

We used machine learning prediction analysis (RFA) to identify the key risk factors (from a pool of 15) that discriminated the two classes of persons identified in the previous analyses. The pool of predictors was selected on the basis of considerable research previously showing independent associations with EF performance, differences or change in aging (e.g., Blasko et al., 2014; Chen et al., 2019; Perkovic et al., 2018; Qiao et al., 2020; Sapkota et al., 2017; Thibeau et al., 2016). Although most of these studies considered one (or few) predictors, we assembled and tested a diverse pool of predictors in a computationally competitive context to identify the most important or significant factors. We observed that eight of the 15 risk factors from four domains (demographic, lifestyle, functional, and mobility) predicted Class 2 membership (see Figure 3). Specifically, the predictors of the unique higher-level and multidimensional class were, in order of importance: younger age, more everyday novel cognitive activity, higher education, lower BMI, lower PP, female sex, faster balance, and more everyday physical activity. Notably, the RFA considers both the independent contributions of each predictor, as well as its participation in interactions with other predictors (Sapkota et al., 2018). The RFA does not identify potential mechanisms or particular interactions by which these predictors may operate. Moreover, the two classes we are comparing have not previously been detected or studied in terms of discriminating biomarkers. Therefore, the present discussion is limited to noting previous results in which the present predictors were found to be independently associated with internal indicators of EF performance level. The identified predictors in the present analyses were selected in the context of a larger pool of potential predictors which are typically examined independently or in small clusters. For both our analyses and single-variable approaches, the identified predictors may be proxies for other unmeasured features not included in the analyses. We encourage further investigation of multiple predictors, their potential interactions, and associated mechanisms.

Demographic

For this domain, the three important factor predictors that discriminated the higher performing and multidimensional class were younger age, higher education, and female sex. Regarding age, research has reported that cortical thinning and volumetric loss in EF-related brain regions (e.g., prefrontal, lateral, and medial cortices) typically occur with advancing age, posing a risk for disruption of EF-related activity in the brain (Li et al., 2017; Yuan & Raz, 2014). Among cognitively normal older adults, relatively younger persons may retain complex EF structure and be less susceptible to performance deficits through lower rates of cortical thinning and volumetric loss in brain regions associated with EF. Education (fewer years of schooling) is an established risk factor for dementia (Livingston et al., 2017), but has mixed associations with non-demented cognitive change (Dixon & Lachman, 2019). As an important predictor in this study, education may operate in support of EF structure and performance through the mechanism of retained cognitive and brain reserve (Stern, 2012). Although the evidence is indirect and mixed, higher levels of schooling among non-impaired older adults has been associated with better EF performance than older adults with lower education—a pattern often attributed to preserved cognitive and brain reserve (e.g., Chen et al., 2019; Tucker-Drob et al., 2009; Ward et al., 2015). Regarding sex (male or female), our results indicated a female advantage in that higher EF performance was associated with females and lower EF performance with males. Although different in approach, some recent EF and aging research has produced related results: (a) women have higher EF performance and significant less decline than men and (b) women improve better than men after practice on inhibitory tasks (Mansouri et al., 2016; McCarrey et al., 2016; Zaninotto et al., 2018). Conceivably, the currently observed advantage may be associated with selective evidence that older females may have greater cortical thickness, lower age-related reductions in frontal and temporal brain volumes, and lower rates of cortical thinning than males (Leonard et al., 2008; Lotze et al., 2019; Pacheco et al., 2015; Shaw et al., 2016).

Lifestyle

The two important predictors from this risk domain were everyday novel cognitive activity (self-report engagement in activities such as reading or playing bridge) and everyday physical activity (self-report engagement in activities such as jogging or walking). Both of these leisure-time activities are associated with significant reductions in the risk of executive dysfunction and even dementia (Cheng, 2016). For instance, research indicates that more engagement in cognitive (Chambon & Alescio-Lautier, 2019; Tang et al., 2018) and physical (Erickson et al., 2015; Thibeau et al., 2019) activity are associated with sustained EF performance in non-demented aging. Contrarily, high frequent engagement in passive activities (e.g., TV viewing) involving low levels of cognitive or physical activity not only correlate significantly but predict lower EF performance (Blasko et al., 2014). In relation to potential mechanisms through which these predictors may work to produce complex EF structure and sustained EF performance in aging, research suggests: (a) cognitive activity strengthens plasticity and function in neural circuits and (b) physical activity preserves brain volume and the structural integrity of neurons (Cheng, 2016; Dixon & Lachman, 2019; Hertzog et al., 2008).

Functional

Two indicators from this risk domain discriminated classes of EF performance and structure: BMI and PP. In general, a higher-than-normal BMI is a risk for pathophysiological changes in vascular health, systemic inflammation, impaired insulin regulation, and poor cardiovascular fitness (Anstey et al., 2011; Arvanitakis et al., 2018; Gunstad et al., 2007). Recent research reported that higher-than-normal BMI in early adult life is associated with lower EF performance in late life (Zhou et al., 2020). Moreover, maintaining optimal BMI levels predicted higher EF performance and sustained trajectories over time (Bohn et al., 2020). Regarding PP as a predictor, lower values of PP (indicating better vascular health in the form of reduced arterial stiffness) have beneficial effects such as preventing hypertension (e.g., mini-infarcts and cerebral vascular damage (Cooper et al., 2016), and reducing AD-related pathophysiology and other neurodegenerative processes (Nation et al., 2013; Warsch & Wright, 2010; Wong et al., 2019). Ultimately, better vascular health could promote positive EF effects with aging, such as the current retained complex structure and higher performance (Caballero et al., 2021; Mason et al., 2020; McFall et al., 2014).

Mobility

Of the two tested mobility indicators, faster balance (measured as timed turn) predicted membership in the higher EF class. Although not well understood in terms of potential mechanisms, it is conceivable that faster balance is a more cognitively complex mobility measure compared to gait (measured as timed walk)—and perhaps more sensitive predictor of the higher EF class detected in this study. Some evidence suggests that faster balance could be associated with higher EF performance across typical EF facets. For instance, studies have shown that individuals with faster balance have better response inhibition in tasks related to everyday activities (i.e., walking; Coppin et al., 2006; Giladi et al., 2007; Kearney et al., 2013; Li et al., 2018). In addition, older adults with faster balance have faster responses and make fewer errors on EF tasks that tap inhibition, shifting, and updating compared to older adults with slower balance (Thibeau et al., 2019; Zhang et al., 2020).

Non-Significant Factors

Despite their previous relevance to EF and aging structure and change (Mirelman et al., 2012; Qiao et al., 2020; Sapkota et al., 2017; Sternäng et al., 2015; Zhang et al., 2013), seven factors did not emerge as predictive of the present classes in the quantitatively competitive context. These were representative of three domains: genetics (IDE, COMT, BDNF, APOE), mobility (gait), and functional (PEF, grip strength). That several factors previously related to EF aging did not perform as important predictors in this machine learning context has several implications. First, it is possible that the criterion of prediction in this study (membership in two novel EF classes) provided a target not directly related to prevailing EF outcomes (Caballero et al., 2021; Sapkota et al., 2017). Second, given the multi-factorial nature of EF aging (performance and structure, as well as variability over time), additional studies should consider not only independent predictor effects but also interactive or synergistic effects among multiple predictors to selectively discriminate differences in EF performance and structure. Third, the quantitatively competitive context of the RFA may have useful applications to investigating broader EF phenomena in aging, such as prediction to unidimensional structure or impaired status. Fourth, several of the predictors of the current classes were potentially modifiable risk factors (Anstey et al, 2020; Dixon & Lachman, 2019; Livingston et al, 2017). These include: education, everyday novel cognitive engagement, physical activity, BMI, PP, and balance. As has been promoted previously (e.g., Livingston et al., 2017; McFall et al., 2019; McDermott et al., 2017) such factors may provide targets for interventions designed to promote healthier EF in aging through risk reduction and protection enhancement.

Limitations and Strengths

Several limitations to the present study should be noted. First, as a group of older adults, our sample is relatively healthy, generally well-educated and cognitively intact. As such, the sample does not fully represent the global population of aging adults. We note especially the lack of racial and ethnic diversity in this age group for both of these research samples and the source population. As we targeted a non-impaired asymptomatic aging group, the present sample may be relatively advantaged in terms of cognition and education. Although this sample arguably represents a large segment of older adults in western industrialized countries, further study in more diverse samples is recommended. Second, both the classes we obtained and the predictors we examined were based on single-wave of EF performance. A longitudinal approach aimed at predicting actual EF performance and structure changes and whether they are coupled with predictor changes would yield supplemental results. Specifically, this approach could test if (a) EF performance remains stable or declines in a second or third wave of measurement, (b) the structure of EF remains multidimensional or is reduced to a unidimensional structure in Class 2, and (c) the same significant predictors discriminate classes of EF performance and structure at different waves of measurement. The present study was designed to focus on the application of data-driven techniques to large single-occasion data in order to accomplish classification and predictions of EF performance and structure. Third, as noted in the statistical analysis section, a separate dataset for external validation was not available in this study. However, we implemented procedures for internal validation using the out-of-bag error in RFA. Future research can consider using independent datasets to assess and evaluate this or related predictive models. Fourth, although our pool of potential risk factor predictors was fairly broad and representative of previous research, not all potential predictors were available (Dixon & Lachman, 2019; Livingston et al., 2017; Stern, 2012). For example, cholesterol level, imaging data and additional genotypes were not available in the current VLS battery. In addition, some risk factors were available in the VLS dataset (e.g., smoking, alcohol) but were not included in the analyses because some categories were not sufficiently represented. For instance, most participants (> 90%) never smoked or have quit and most (87 %) currently consume alcohol (although lightly). Future research can consider implementing and testing more (and an even broader range of) predictors. Fifth, after the classes were established in the first goal, we observed that Class 1 had a higher mean age than Class 2. The group difference was relatively small in magnitude and the two classes were similar in age range. It is possible that this age difference could have influenced some aging-related predictors. However, RFA limits the amount of predictor variables used in each decision tree (Strobl et al., 2009), giving the opportunity for non-aging related predictors to also exert their influence on the RFA model; this could have occurred independently of the effects of aging-related predictors.

Regarding strengths, we first note that the use of a large, well-characterized sample of individuals distributed across a broad range of aging was essential to capture the differentiated patterns of performance observed in the two classes. Second, the latent classes were detected by employing a new analytic approach to examining individualized patterns of EF performance and structure. Specifically, we used FMM as a person-centered approach, which allowed us to evaluate the relationship between EF performance and structure. This approach led to the detection of a novel class of EF aging characterized by higher EF performance in a multidimensional EF structure. Third, we used RFA, a powerful machine learning technology for evaluating the relative importance of multiple risk factor predictors (Couronné et al., 2018; McFall et al., 2019; Sapkota et al., 2018). RFA enabled us to simultaneously compare multiple predictors (previously independently related to EF aging), determine the most important exemplars, and produce an order of predictor importance within this dataset. Notably, the present predictors spanned a continuum of risk domains for cognitive decline, impairment, and dementia.

Conclusion

A series of data-driven, person-centered analyses provided novel information about EF performance and structure in aging. Specifically, we observed that individual differences in EF aging can be characterized by two discriminable clusters: (a) lower EF performance and unidimensional structure (Class 1) and (b) higher EF performance and multidimensional structure (Class 2). For existing theoretical approaches to EF and aging, the results provided a complementary perspective in which multidimensionality can be associated with higher performance levels. Given widespread interest in successful brain and cognitive aging, we next examined a notable number of factors (n = 15) associated with EF aging, comparing them simultaneously in a quantitatively competitive context. Notably, membership in Class 2 was predicted by a set of eight risk/protective factors representing four domains: demographic (age, education, sex), lifestyle (everyday novel cognitive activity, everyday physical activity), functional (body mass index, pulse pressure), and mobility (balance). For all factors, the direction of prediction was mechanistically sensible (e.g., predicting membership in the higher class were younger age, higher education, more cognitive activity, lower pulse pressure)—although mechanisms per se were not examined. That eight predictors emerged as selectively predictive of the class membership implies that seven potential predictors were not significant in the analyses. These included genetic factors, previously associated independently in studies of EF aging. In sum, EF performance and structure in aging may be more nuanced than previously observed. Detectable and classifiable different patterns were predicted by a selective network of markers representing multiple modalities of risk and protection. Many of these markers (e.g., everyday novel cognitive activity) are modifiable and therefore may provide targets for early interventions that could promote sustained individualized membership in the class characterized by higher performance and multidimensional structure of EF in aging.

Key Points.

Question:

What do data-driven and person-centered modeling approaches to executive function (EF) aging reveal about structural classes and differential risk predictors?

Findings:

Our modeling approach revealed the existence of two EF classes, the commonly observed unidimensional and lower performing subgroup and the novel multidimensional and higher performing subgroup. These subgroups are discriminated by a selective set of brain aging risk factor predictors.

Importance:

In non-demented aging, EF performance and structure characteristics are more nuanced than previously observed but predicted in part by modifiable risk markers that could be targeted for interventions to promote healthier EF aging.

Next Steps:

Apply longitudinal versions of these data-driven modeling technologies to examine the progression of performance-structural change and determine the key biomarkers predicting different trajectories.