This study identified distinct cognitive dimensions that seem to be subserved by shared and dissociable structural brain patterns in older age. Older adults with impaired cognitive flexibility, due to rigid focusing and reduced feedback sensitivity, show reduced GMV in localised regions. In contrast, those with slow response tendencies and heightened non-decision times were susceptible to increased brain aging relative to chronological age. A single factor composed of SES and cognitive reserve variables was significantly predictive of cognitive profiles. Our work highlights potential avenues for remedying age-related cognitive health disparities in a Global South context, namely by using SES and cognitive reserve variables as modifiable targets for intervention.
Methods
Recruitment Strategy
A proportion of participants were recruited via government-linked community centres for senior citizens86, advertisements in social media and newspapers, and through referrals of existing participants. Others were recruited from low-cost housing areas (Projek Perumahan Rakyat87 or People’s Housing Project). Through liaising with a charity that regularly provides food parcels to these housing areas, we arranged for mass on-site recruitment and screening sessions with communities. Stratified sampling was conducted where we recruited participants of Malay, Chinese, Indian, and other minority ethnic backgrounds, as well as ensured we had sizeable numbers of low and middle-to-high income participants. This study received approval from Sunway University’s Research Ethics Committee (SUREC2020/039).
Inclusion and Exclusion Criteria
Participants had corrected or normal-to-corrected vision and hearing, were generally mobile (able to walk at least 3 metres unassisted), were able to communicate in at least one of the following languages: English, Malay, or Mandarin, and did not have confirmed diagnoses of neurodegenerative and/or psychiatric disorders and a history of stroke. Standard MRI contraindications were also evaluated.
Additionally, we ruled out severe cognitive impairment or probable Alzheimer’s disease using the Montreal Cognitive Assessment (MOCA9). We defined the MOCA cut-off to a score of 13 for our study for 3 reasons: 1) The original cut-off of 269 for MOCA is criticised to be too high, even in samples of highly-educated older adults88; 2) the MOCA is highly sensitive to one’s education level across various cultures35,89–91, and many of our participants living in poverty reported having little to no education. Using the original cut-off, or even the cut-off of 18 as recommended in a study reporting results in a Malay-speaking sample92, would have led to the exclusion of a significant proportion of low-income participants; 3) several papers recommend a cut-off of 13-to-14 in less-educated or ethnic minority groups for ruling out potential Alzheimer’s disease91,93–95.
Questionnaires
Socioeconomic Status (SES)
Our questionnaire included several questions probing socioeconomic standing. First, participants’ highest education level was measured on a 7-point scale: 0 (No education), 1 (Primary school), 2 (Secondary school), 3 (Pre-university), 4 (Diploma – equivalent to the 1st year of a Bachelor’s degree), 5 (Bachelor’s degree), 6 (Master’s degree), and 7 (Doctor of Philosophy). Next, participants’ occupational complexity was determined by coding participants’ longest held occupation on a scale from 1–5 (1 – least cognitive difficulty; 5 – highest cognitive difficulty) following an established convention96. We also determined whether participants were English speaking (yes/no) based on reports of English proficiency being indicative of quality of local education97, availability of school funding/resources97, and considered an important factor for employability98 in Malaysia.
Objective SES was determined based on participants’ current housing (low-cost or not), a crowding index (number of people in household/number of bedrooms), and whether they were receiving any government financial support. Subjective current SES (i.e., perception of one’s social class) was evaluated using the MacArthur scale of subjective social status, a pictoral scale that uses a symbolic ladder with rungs labelled from 1–10. Participants were asked to identify their perceived place in society (either the top, middle, or bottom of the ladder) using the numbered rungs.
Next, food security as influenced by financial ability99 was probed using two questions: 1) “Currently, how often do you or anyone in your household skip meals, or reduce their portion because there is insufficient money for food?”; 2) “Currently, how often are you (or your household) short of money to buy enough food?” Participants’ responses ranged from 1 (Never) to 7 (Very Often). Scores per question were summed to obtain an aggregate score (range: 2–14), with higher scores indicating greater food insecurity.
We also sought to explore SES from a life course perspective100, by evaluating how early and later life exposures independently or accumulatively impact geriatric cognition. Life stages were divided into childhood-adolescence (6–19 years), middle adulthood (20–59 years), and late adulthood (60 years to current age). For each life stage, we asked participants to rate the financial situation of their household (1: Extreme poverty to 7: Extremely well-off) and to rate the financial situation of their household in comparison to other families (1: A lot worse off, 7: A lot better off).
Physical Health
We requested that participants disclose whether they had any diagnosed health issues, with specific questions probing diabetes, high cholesterol, and hypertension/high blood pressure, which are among the most prevalent health conditions observed in Malaysian older adults101.
Physical activity frequency was measured using a question from the Outdoor Health Questionnaire developed by the National Institute for Health and Care Excellence: “In the past week, on how many days have you accumulated at least 30 minutes of moderate intensity physical activity such as brisk walking, cycling, sport, exercise, and active recreation? (Do not include physical activity that may be part of your job or usual role activities.)”. Participants could answer between 0 and 7 days.
Any recent reductions in food intake were also noted using a question from the Mini Nutritional Assessment-Short form102 (“Has food intake declined over the past 3 months due to loss of appetite, digestive problems, or chewing or swallowing difficulties?” 0: Severe decrease in food intake, 1: Moderate decrease in food intake, 2: No decrease in food intake).
Mental Health and Sleep
The Centre for Epidemiologic Studies-Depression Scale (CES-D) with 14-items103 was used to assess symptoms of depression. Participants rated items such as “I felt depressed” on a rating scale from 0 (rarely or none of the time) to 3 (most or all of the time). The 6-item short form of the Spielberger State-Trait Anxiety Inventory (STAI-S)104 was used to measure current anxiety symptoms, where participants rated items such as “I feel tense” on a scale from 0 (Not at all) to 3 (Very much).
To assess subjective well-being, the Satisfaction with Life-Scale105 was administered, consisting of 5 statements that participants have to rate their agreement on (1: Strongly Disagree – 7: Strongly Agree).
Next, to account for the potential unique impact of financial struggles on mental health symptoms, we asked participants the following questions: 1) Currently, how often do you worry that you might not be able to pay back your debt(s)?; 2) Currently, how often do you worry that you (or your household) might not have enough money to make ends meet?; 3) Currently, how often do you worry that there wouldn’t be enough money to buy food?; 4) Currently, how often do you worry that you (or your household) might lose your home? Participants rated the statements on a 7-point scale (1: Never – 7: Very Often). An aggregate score was obtained ranging from 4 to 28, with a higher score indicating greater financial worries. Similar questions have been utilised previously utilised to research financial anxiety106,107.
Sleep duration and quality were assessed by asking participants the following questions from the Pittsburgh Sleep Quality Index108: “During the past month, how many hours of actual sleep do you get at night? (This may be different than the number of hours you spend in bed)” and “During the past month, how would you rate your sleep quality overall?”. Participants rated the latter question on a 4-point scale (1: very bad, 2: fairly bad, 3: fairly good, 4: very good).
Lifestyle and Cognitive Engagement Measures
Lastly, our study also considered various lifestyle variables comprising social and cognitive engagement, and leisure. For social engagement, participants were asked to select the extent to which they had been socially active over the past 1 month out of 3 options: 1) “I am socially active, including meeting family, friends, and new people”, 2) “I am socially active, but with family members and close friends only” 3) “I prefer to stay at home or keep myself busy with other activities.”
As internet use is reportedly associated with improved cognitive functioning in older adults109 we included a question probing how often participants use the internet in a given week (0 to 7).
The cognitive engagement measures included, first, a question related to book ownership adapted from a prior study110: “About how many books are in your home?” followed by the response options “few (0–10)”, “enough to fill one shelf (11–25)”, “Enough to fill one bookcase (26–100)”, “enough to fill two bookcases (101–200)”, “enough to fill more than two bookcases (more than 200)”.
Next, we assessed participants’ reading habits using two questions: “Currently how often do you a) read books, b) read the news. Participation in other cognitively stimulating activities111 aside from reading was measured using 3 questions: “Currently how often do you a) play tabletop/board games, b) attend art classes and c) attend hobby-based classes (writing, singing, etc). Participation in leisure activities112,113 was assessed using the following 4 questions: “Currently how often do a) listen to the radio, b) watch TV and/or videos, c) attend live performances, d) Go to the cinema or watch movies”. Participants rated each question on a scale from 1 (Never) to 7 (Very Often).
We adopted a life course approach here motivated by prior research reporting that engagement in physical and cognitive activities across the lifespan can significantly impact later-life cognition114,115. Hence, participants were requested to also answer questions related to physical activity, reading, cognitive activities, and leisure activities but recalling a typical week in their mid-adulthood (20–59 years) and childhood (6–19 years).
The full questionnaire was administered to participants in either English, Malay, or Chinese languages (depending on participant preference). When validated translations of scales were not available, we applied the translation procedure recommended by the Psychological Science Accelerator116, whereby the English version of the scale was translated by two native speakers followed by back translations by a 3rd native speaker. Reliability metrics for aggregate scores for each language are reported in Table S7.
Cognitive Tasks
Before beginning the computerised neuropsychological tasks, participants were asked whether they were comfortable using a computer mouse. Those who were unaccustomed first completed a series of mouse practice games (“Mousercise”) to gain familiarity with using the device. The two tasks described below were completed alongside other cognitive tests not reported here.
WCST
The WCST from the Psychology Experiment Building Language programme117 was presented on a laptop to participants. The task contains up to 128 trials. Participants were shown 4 decks on-screen with a different combination of colours, numbers, and shapes. They were instructed to sort cards appearing at the bottom of the screen, using a computer mouse, according to one of three rules at a time, either number, colour, or shape. The rule must be discovered using trial and error via visual feedback received after each card is sorted. Cards were sorted by clicking on the chosen deck using the laptop mousepad. If a card is sorted correctly, the feedback shown would be ‘Correct’. If the card is sorted incorrectly, the feedback shown would be ‘Incorrect’. There was no time limit for a card to be sorted on each trial, but participants were told to answer as quickly and as accurately as possible.
After 10 cards have been successfully sorted consecutively, one set is completed and the sorting rule changes. The process continues until the participant either sorts all 128 cards or they complete 9 sets.
The following behavioural measures were obtained: proportion of correct responses, proportion of perseverative errors (incorrectly choosing a deck based on the rule from the previous set),, proportion of non-perseverative errors (errors that were not perseverative), proportion of unique errors (where a deck is chosen that does not match the test card on any rule), number of trials needed to complete first set (out of 128), and number of sets completed (out of 9).
Go/No-Go Task
The Go/No-Go task was programmed and presented using Psychopy118 and the design was informed by prior research119–121. On each trial, participants were shown the letter ‘M’ or ‘W’ in the centre of a computer monitor screen. Half of the participants were instructed to make a ‘Go’ response (via pressing the Spacebar button) when they saw ‘M’ and to withhold responding when they saw ‘W’; the remaining participants completed a version in which “W” was the Go stimulus and “M” was the No–Go stimulus. Assignment to either version of the task was based on participant ID number, wherein even-numbered IDs had “M” as the Go stimulus, while odd-numbered IDs had “W” as the Go stimulus. Responses were registered on a computer keyboard placed in front of participants on a table. Each trial began with a fixation point presented for 500 milliseconds (ms) followed by either Go or No-Go stimulus for 100ms, and finally a blank screen. Participants were instructed to respond within 500ms of stimulus onset. Feedback showing ‘Too slow!’ appeared after responses exceeding this timing, and ‘Incorrect’ was shown after any erroneous responses to the No-Go stimulus. No feedback was given following correct responses. The task comprised 500 trials (80% Go and 20% No-Go stimuli). Participants received a two-minute break halfway through the task.
We calculated the following behavioural measures: proportion of Go errors (incorrectly not responding to Go stimulus), proportion of No-Go errors (erroneously responding to No-Go stimulus), Go reaction times (RT, in seconds [s]) and No-Go RTs.
MRI Protocol
Due to limitations in number of research participants hospital radiology units were willing to accommodate, imaging data was collected across two sites, namely Sunway Medical Centre (SunMed) and Sunway Medical Centre Velocity (SMCV), both located in Klang Valley, Malaysia. T1-weighted MPRAGE images were acquired from all participants. In SunMed, participants were scanned using a Siemens Magnetom Skyra 3-tesla (3T) scanner with a 20-channel head coil and the following parameters: repetition time (TR) = 2110 ms, echo time (TE) = 3.52 ms, field of view (FOV) = 240 x 240 mm, slice thickness = 0.94 mm, flip angle = 9°, voxel size = 0.9 x 0.9 x 0.9 mm³. In SMCV, a Signa Pioneer 3T scanner with a 21-channel head coil and the following parameters: TR = 8.3 ms, TE = 3.40 ms, FOV = 240 x 240 mm, slice thickness = 1.00 mm, flip angle = 9°, voxel size = 0.5 x 0.5 x 1.0 mm³, was used. Given the heterogeniety in scanning parameters between sites, we controlled for ‘site’ in all MRI analyses.
Data processing was performed using SPM12 release 6685 (Wellcome Department of Imaging Neuroscience, London, UK1) implemented in MATLAB R2022b (Mathworks Inc., Natick, MA, USA). All data were manually reoriented and co-registered to the Montreal Neurological Institute 152 (MNI152) template.
Data were then processed using voxel-based morphometry, specifically segmentation and normalisation were conducted via the Diffeomorphic Anatomical Registration, using the Exponential Lie algebra algorithm (DARTEL) technique122. Default parameters were used in SPM during segmentation except that affine regularisation was performed using the template for East Asian, instead of European, brains. Grey matter (GM), white matter (WM), and cerebrospinal fluid (CSF) tissue maps were used to create a custom template based on the larger sample of 397 participants. For each participant, flow-fields were computed during template creation to provide the transformation matrix from each native image to the template. Normalisation was applied to align and warp each participant’s data to the DARTEL-created template, while smoothing was conducted using an 8-mm full width at half maximum (FWHM) Gaussian kernel to remove potential artifacts and account for imperfect alignment of brain data from different participants. We computed the total intracranial volume (ITV) from GM, WM, and CSF tissues obtained during the segmentation step as a metric to be included as a nuisance regressor in later regression models to control for individual differences in overall ITV.
Imaging data were subjected to both whole-brain and region-of-interest (ROI) analyses. To understand brain-Go/No-Go performance relationships we defined the following ROIs: bilateral insula56, inferior frontal gyrus56, supplementary motor area (SMA)57,123. For the WCST, the ROIs comprised the bilateral prefrontal cortex (PFC)12 and dorsolateral (PFC)124. We also defined the following ROIs: bilateral hippocampus65, amygdala4, and cerebellum3, based on previous literature reporting reduced GMV in these regions being associated with poor cognition in aging populations. All ROIs were selected a priori. Bilateral anatomical masks for each region were created in WFU-PickAtlas125. The Talairach Daemon atlas was used to create a cerebellum mask, the Automated Anatomical Labelling atlas was used to create a SMA mask, while the Brodmann atlas was used to create masks for all other ROIs.
BrainAge
We applied a pre-trained machine learning pipeline (brainageR) to predict the ages of our participants30. BrainegeR was utilised as it consistently outperforms other brain-age models on accuracy and test-retest reliability126,127. Using SPM12, brainageR first segments the T1 scans into GM, WM, and CSF probability maps and then conducts normalisation. Subsequently, in R language, these normalized probability maps are concatenated into one vector and principal component analysis (PCA) is conducted. A Gaussian process regression model is then used to predict brain-age using the first 435 components from the PCA. The model was previously trained on a sample of 3377 healthy individuals (range 18–92 years) from seven publicly available datasets.
Each participant’s age was subtracted from their predicted brain-age (BrainAge – ActualAge) and then inserted as a predictor variable in regression models, whereby more positive values signify advanced brain-age relative to chronological age.
Statistical Analysis
WCST Reinforcement Learning Model
A reinforcement sequential learning model128 was fitted to trial-by-trial WCST data.
The full model contained 4 free parameters, namely the reward rate (r, how quickly attention weights change to rewarding feedback), punishment rate (p, how quickly attention weights change to punishing feedback), decision consistency (d, how much deck choice is influenced by attention weights), attentional focusing (f, only important on trials with ambiguous feedback and represents the degree to which the update is focused only on the category/rule with the largest attention weight).
The dependent variables put into the model are 1) an outcome variable which represents whether a trial was rewarded or not (1 or 0) and 2) a matching matrix which quantifies which categories (colour, number, or shape) associated with a chosen deck match the test card. For instance, if the chosen deck matched the test card based on colour but not number or shape, the matching matrix, m, for that trial would be defined as
m = \(\left[\begin{array}{c} 1 \\ 0\\ 0\end{array}\right]\)– Eq. 1
The model calculates the probabilities associated with choosing each deck as a function of attention weights (a), which represents the weight/value given to each category per trial. The matrix elements of the attention signal always sum to one. It was assumed that for each participant’s first trial, the attention weights are divided evenly between categories:
a = \(\left[\begin{array}{c} .333.. \\ .333..\\ .333..\end{array}\right]\)– Eq. 2
Attention weights are updated using a feedback signal (s), which represents whether the categories were rewarded or not. For example, in the case where a chosen deck matches the test card based on only colour and nothing else, and the trial is rewarded, the feedback signal would look like this,
s = \(\left[\begin{array}{c} 1 \\ 0\\ 0\end{array}\right]\)– Eq. 3
with each element of s representing colour, number, and shape respectively. The current attention weights are updated based on the feedback signal according to the following equations:
a t+1 |rewardedt = (1-r)at + rs - if trial was rewarded – Eq. 4
a t+1 |punishedt = (1-p)at + ps - if trial was punished – Eq. 5
where t refers to the current trial. Parameters r and p determine how rapidly attention weights alter based on rewarding and punishing feedback signals respectively.
In the example above, the feedback is unambiguous as the chosen deck matches the test card on only one category. However, in some cases where more than one category is matched (for example, both colour and shape), the feedback signal relies on the free parameter f to modulate how focused or wide the attention is for each category’s feedback. When f approaches 0, attention is split evenly between the matching categories:
s = \(\left[\begin{array}{c} 0.5 \\ 0.5\\ 0\end{array}\right]\)– Eq. 6
As f increases, the feedback signal is split proportionally between current attention weights. For example, if the attention weight for colour is higher than shape, the feedback signal would follow suit and perhaps be represented by:
s = \(\left[\begin{array}{c} 0.75 \\ 0.25\\ 0\end{array}\right]\)– Eq. 7
The following equations represent how the feedback signal is modulated by the attention weights and the matching matrix.
st|reward = \(\frac{{\mathbf{m}}_{\text{t}}{\mathbf{a}}_{\text{t}}^{f}}{\sum {\mathbf{m}}_{\mathbf{t}}{\mathbf{a}}_{\text{t}}^{f}}\) – Eq. 8
st|punish = \(\frac{(1-{\mathbf{m}}_{\text{t}}){\mathbf{a}}_{\text{t}}^{f}}{\sum (1-{\mathbf{m}}_{\text{t}}){\mathbf{a}}_{\text{t}}^{f}}\) – Eq. 9
When outcome on the current trial is correct, the feedback signal is computed only with the matching attention weights, and when the outcome is incorrect, only the non-matching attention weights contribute to the feedback signal.
Finally, the probability of choosing a specific deck is defined as
P = \(\frac{{\mathbf{m}}_{t}^{{\prime }}{\varvec{a}}_{t}^{d}}{\sum {\varvec{a}}_{t}^{{d}^{ }}}\) – Eq. 10
where the d parameter influences the predicted probability of choosing a deck per trial. As d becomes lower, choices become more random and less dependent on attention weights (more exploratory). As d becomes higher, choices are heavily constrained by attention weights (more exploitative). \({\mathbf{m}}_{t}^{{\prime }}\) is simply the matching matrix, \({\mathbf{m}}_{t}^{ }\),transposed to enable matrix multiplication (dot product) with \({\varvec{a}}_{t}^{ }\).
The full model was compared to 4 degenerate models. Each degenerate model had one parameter fixed to assess the contribution of each parameter to capturing behaviour on the task.
The 1st alternative model (rpd0) fixed the f parameter to be 0, the 2nd model (rpd1) fixed f to be 1, the 3rd (rp1f) fixed d to be 1, and the 4th alternative model (rrdf) assumed a single common learning rate for both reward and punishment.
Model code was adapted from a prior study12.
We assumed Beta (0,1) distributions as priors for all free parameters12. Parameters r and p naturally varied between 0 and 1 while d and f were rescaled to range between 0 and 5.
Computation of posteriors per parameter were conducted using a Bayesian approach via Markov Chain Monte Carlo (MCMC) sampling implemented in JAGS129 software. Four randomly initialised MCMC chains were run during model-fitting. Convergence of chains was confirmed using the potential scale reduction statistic R̂. A cut-off R̂ value of 1.2130 was used to check that the chains were well-mixed for each parameter.
We evaluated model fit using the Deviance Information Criterion (DIC), which is typically used in Bayesian model selection, and takes accuracy and model complexity (number of free parameters) into account131. We also computed the penalised deviance132 which penalises model complexity more stringently than standard DIC. The lower the DIC and penalised DIC, the better the model fit.
Go/No-Go Drift Diffusion Model (DDM)
DDM assumes that fast two-choice decisions (in the case of Go/No-Go, the decision to either respond or withhold responding) are made via a stochastic process where information is accumulated over time. After a stimulus is encoded, the point at which the decision process begins is determined via a starting point parameter (z) wherein higher z values indicate preponderance for responding (Go) while lower z values reflect a bias for not responding (No-Go). From the starting point, the decision process drifts (in random walk motion) between the upper boundary (respond) and a lower implicit boundary (withhold responding, denoted by 0). The rate at which a participant accumulates evidence to reach a decision is determined by the drift rate v. The better the information quality or the more efficient one’s ability to accumulate evidence, the greater the values of v. When the decision process reaches either one of the boundaries, a decision is finally made. The distance between upper and lower boundaries is denoted by a boundary separation parameter (a) where larger a-values indicate higher emphasis on accuracy while smaller a-values reflect valuing speed over accuracy. Reaction times (RT) from trial-to-trial are assumed to be determined by the sum of the decision times and non-decision times. The non-decision time parameter (Ter) stipulates time taken up by processes peripheral to decision-making including stimulus encoding and execution of a motor response.
Four drift diffusion models were fitted to trial-by-trial Go/No-Go choice and RT data using the Hierarchical Bayesian estimation of the Drift Diffusion Model133 (HDDM) toolbox. Trials with reaction times below 0.1s were removed before model-fitting121. RTs during non-responses for both Go and No-Go trials were coded as not a number (NaN).
Model 1 assumed separate drift rates for Go (v.go) and No-Go (v.nogo) trials, as well as z, a, and Ter parameters that were free to vary. Model 2 was akin to the first model except z was fixed at 0.5 indicating no bias in starting point between upper and lower boundaries. Model 3 was also like Model 1 but with only a single drift rate (v) capturing evidence accumulation across all Go and No-Go trials. Model 4 was like Model 3 except that z was fixed to 0.5. Such variations of models have been explored and compared in prior research134. For models with separate drift rates for Go and No-Go trials, HDDM output a drift scaling parameter (dc) which was subtracted from v per participant to obtain v.nogo, and added to v per participant to obtain v.go.
Model-fitting was conducted using G square approach (optimisation set to ‘gsquare’ in HDDM), based on chi-square optimisation, where the Go/No-Go data are fitted based on RT quantiles13,14. The RT distributions for responses to Go and No-Go stimuli were represented by the 0.1, 0.3, 0.5, 0.7 and 0.9 quantiles, while non-responses were represented by a single bin as there are no observable RTs. We assumed 5% outliers81 in our data which were modelled under a different process (outliers in HDDM are modelled using a uniform distribution133). Model-fitting using Gsquare was conducted following a freely available tutorial (“Fitting go/no-go using the chi-square approach”: http://ski.clps.brown.edu/hddm_docs/demo_gonogo.html) and has been applied in prior work81,135.
The Bayesian Information Criterion (BIC), which can be readily calculated in HDDM, was used to select the best-fitting model (the lower the BIC value the better the fit). The BIC penalises model complexity better than other model-comparison metrics (e.g., Akaike Information Criterion)136.
Principle Component Analysis (PCA) of cognitive data
We obtained profiles of cognitive performance separately for WCST and Go/No-Go data using a data-driven approach. All performance measures across participants were z-scored and subjected to an exploratory PCA using the psych (ref) library in R. Advantages of utilising PCA are that the technique decomposes the original data into corresponding principal components (PCs) that are orthogonal to each other and enables capture of independent features of the data. We included both standard (model-free) performance measures alongside computational model parameters (see correlations in Tables S8.1 and S8.2) in the PCAs to facilitate interpretability of resulting cognitive profiles by elucidating how different measures load onto each other. In fact, this method has been used in prior research to guide interpretation of computational model parameters across different reinforcement learning tasks137. PCs were extracted based on visual inspection of the screeplot and the Kaiser’s criterion (eigenvalue > 1). From these results, a 3-factor structure was used for Go/No-Go and a 4-factor structure for WCST. Promax rotation was used for both PCAs. Resulting loading patterns were inspected in each PC and interpreted. Scores per participant for each PC were extracted and used in further analysis.
GMV analysis
To understand how cognitive profiles and computational parameters correspond to grey matter morphology, we entered the profiles and parameters into multiple general linear models (GLM) in SPM12, with whole brain and ROI GMV as dependent variables. Age, gender, ethnicity, MRI scanning site (to account for effects of using multiple scanning sites during acquisition138), and intracranial volume were inserted into the models as covariates of no interest. For whole brain analysis, an absolute implicit mask with a recommended threshold value fixed at p > .02 was used139. Otherwise, for ROI analysis, predefined masks were used (see MRI data acquisition and processing protocol above). To assess the independent effects of cognitive profiles on GMV, we entered the PCs one-by-one into the GLMs. We also obtained whole-brain and ROI maps for contrasts between PCs (e.g., PC1 > PC2). Relationships between GMV and PC/parameter values if they exceeded a voxel-wise threshold of p < 0.05 family-wise error (FWE) corrected, and reached a minimum cluster size of 20 voxels3. Resulting coordinates of brain maps were labelled using a freely available toolbox in R (label4MRI, GitHub - yunshiuan/label4MRI: Label the brain MNI coordinate by AAL/BA system).
Factor Analysis of Survey Variables
Next, we sought to analyse how demographic, SES, health, and lifestyle factors (see Table S1 for complete list) may contribute to variations in cognitive profiles. We also included the verbal fluency test score from the MOCA as verbal fluency or intelligence is reported to be a robust measure of cognitive reserve in older adults140. To prepare the questionnaire data for this analysis, we imputed any missing data using a principal component method called ‘factor analysis for mixed data’ implemented via the ‘imputeFAMD’141 function in R. The method is suitable for mixed (categorical and continuous) data and uses the similarity between individuals and relationship between variables (from conducting factor analysis) to predict missing values141.
Following this, we aimed to a) conduct dimensionality reduction of the data due to the volume of questionnaire items, and b) understand specific constructs that the different questionnaire variables contribute to and how they are interlinked. For these purposes, an exploratory factor analysis with 5 factors (based on Scree plot inspection) using promax rotation was performed on the questionnaire measures. The ‘factanal’ function in R was used. Prior to this, the Kaiser-Meyer-Olkin (KMO) Measure of Sampling Adequacy and Bartlett’s test of Sphericity were conducted to ascertain appropriateness of factor analysis on this dataset. All variables were z-scored before being entered into the factor analysis. A loading score of ± 3.0 was used as a threshold for determining whether a questionnaire variable significantly contributed to a factor142.
Linking Brain and Sociodemographic Factors to Cognitive Profiles
Multiple multivariate GLMs were conducted with the questionnaire factors, Brain Age – Actual Age, and total GMV as predictor variables, and the cognitive profiles as dependent variables (z-scored). All models controlled for gender, age, and ethnicity. Benjamini-Hochberg corrected p-values were computed to correct for multiple comparisons.
To test the robustness of the linear models, we employed ridge regressions with 10-fold cross validation. Moderate correlations and collinearity (determined via variance inflation factor) between predictor variables confirmed that ridge regressions were appropriate. The ridge regressions divided the sample into 10 groups and trained the models on 9/10th of the data and tested the model on the final 1/10th of the data enabling out-of-sample coefficient estimates and R-squared values. In addition, this method performs L2 regularisation where the penalty term λ penalises large coefficients and prevents overfitting. Another advantage is that coefficient estimates are shrunk towards 0 which reduces the impact of irrelevant features on model prediction and enables greater interpretability.
Predictor variables were deemed ‘significantly’ associated with cognitive profiles if GLM-computed p-values and confidence intervals indicated significance, and if coefficients output by the ridge regressions were suitably greater than 0 (coefficient > 0.01).
All tests reported here were two-tailed with p < .05 indicating significance.