Empirical examination of working memory performance and its neural correlates in relation to delay discounting in two large samples

The neurobiological basis of working memory and delay discounting are theorized to overlap, but few studies have empirically examined these relations in large samples. To address this, we investigated the association of neural activation during an fMRI N-Back working memory task with delay discounting area, as well as in-and out-of-scanner working memory measures. These analyses were conducted in two large task fMRI datasets, the Human Connectome Project and the Adolescent Brain Cognitive Development Study. Although in-and out-of-scanner working memory performance were significantly associated with N-back task brain activation regions, contrary to our hypotheses, there were no significant associations between working memory task activation and delay discounting scores. These findings call into question the extent of the neural overlap in delay discounting and working memory and highlight the need for more investigations directly interrogating overlapping and distinct brain regions across cognitive neuroscience tasks.


Introduction
Delayed reward discounting (DRD), also known as temporal discounting, is the cognitive process reflecting preferences for smaller immediate rewards compared to larger delayed rewards (Craft et al., 2022;Downey et al., 2022;Loewenstein et al., 1988) and is considered a psychological measure of impulsive decision-making.Operationally, DRD is J o u r n a l P r e -p r o o f typically assessed by having participants complete a series of forced choices between a smaller amount of money available immediately and a larger amount of money available after a delay.
Across choices, an index can be calculated for each participant's DRD.Level of impulsive DRD is associated with substance misuse, gambling disorder, obesity, and attention-deficit hyperactivity disorder, making the construct central to the study of self-control (Amlung et al., 2017;MacKillop et al., 2011;Matta et al., 2012).Working memory (WM) is a related cognitive construct that refers to the system of short-term memory that is capable of performing cognitive processes on held information (RepovŠ & Baddeley, 2006).Both DRD and WM can be defined within the broad construct of executive function (Buschman & Miller, 2022), or higher-order cognition requiring elements of abstraction, conscious reasoning and mentation, and deliberation.
The association between these two constructs has been given considerable attention in the literature.For example, Wesley and Bickel (2014) published a meta-analysis on the functional overlap of WM and DRD in which they meta-analytically integrated studies that have 449 foci of DRD, 452 foci of WM, 450 foci of finger tapping task, and 450 foci of response inhibition tasks.Subsequently, they performed activation likelihood estimation metaanalyses and noted an apparent overlap between DRD and WM in small clusters in the left middle frontal cortex, anterior insula, and inferior frontal gyrus.A larger cluster of overlap was observed in the left lateral prefrontal cortex.Since this review was published, it has been widely accepted that DRD and WM have a shared neurobiological basis.For example, in the same J o u r n a l P r e -p r o o f year, Finn et al. (2015) investigated the effects of a WM load on DRD in a sample of 623 young adults who varied in degree of externalizing psychopathology.They discovered that the WM load increased discounting in the chosen sample, supporting a strong inverse association between WM and DRD in those with externalizing psychopathology.Moreover, Aranovich et al. (2016) recruited 19 healthy adults to perform an N-back task and complete a DRD questionnaire while undergoing fMRI.They designed mixed block and event-related fMRI experiments with four block types: 0-back, 1-back, 4-back, and a DRD task.It was found that a cognitive challenge, such as a 2-back task, was able to increase the activation in the dorsolateral prefrontal cortex, a region consistently associated with impulse control (Aranovich et al., 2016).This further supported the inverse relationship between DRD and WM.
Additionally, there is evidence in the literature that the WM training is effective in reducing impulsive discounting (Bickel et al., 2011;Finn et al., 2015;Renda et al., 2015).However, no large-scale studies have investigated the neural basis of this overlap between the two cognitive processes within the same dataset.This approach is especially beneficial if the investigated dataset is large and diverse, as it may offer valuable insights into the neurobiological foundations of the overlap between these constructs, directly illuminating the pertinent brain regions, as opposed to other approaches that utilized relatively small sample sizes, as described above.
In the current study, we investigated the associations of brain activations during the fMRI WM task (i.e., N-back) with DRD area under the curve scores in two large neuroimaging datasets, the Human Connectome Project -Young Adult (HCP-YA), which is a brain imaging and behavioural dataset from young adults that is aiming to map the brain's connectivity  al., 2022;Van Essen et al., 2013).Based on the prior literature, our prediction was that WMinduced activations would be inversely associated with the steepness (impulsiveness) of DRD.
As an internal validation of our analyses, we additionally investigated the relationship of the brain activations induced by the N-back task with both the in-scanner behavioural performance and out-of-scanner performance on a separate WM task, predicting that brain activity would be associated with the behaviour on the fMRI task itself and another cognitive task in the same domain.

Human Connectome Project
The HCP includes brain scans (structural and functional) and behavioural data on 1,200 healthy young adults with an age range of 22-35.It was collected between the years 2012 and 2015.In this paper, we used data from 1001 participants of the HCP sample.Table 1 shows the demographics of the investigated sample.The data sample was reduced from 1206 to 1001 as we excluded all the participants with inconsistent, or unavailable data in any of the investigated measures.The inconsistent responses in the DRD data were defined as any response that has more than three inconsistencies, where a single inconsistency was defined as having an indifference point for a given delay that was larger than that of an indifference point for a shorter delay.Additional secondary DRD pipelines were carried out such as the hyperbolic model fit filtration and Jhonson and Bickel's rules (Johnson & Bickel, 2008).Refer to section 2.3.1.on quality control for more details.
For the fMRI images, only subjects with unavailable images were excluded as the quality of the images was maintained and assured by the data releasers.In short, the initial number of participants was 1206, but only 1058 of them had delayed discounting data.After J o u r n a l P r e -p r o o f filtering for consistent delay discounting, 1038 participants remained.This number was further reduced to 1001 individuals who had available N-back fMRI scans.

Adolescent Brain Cognitive Development Study
The ABCD study is an ongoing longitudinal study that has the goal of acquiring brain scans (structural and functional) and behavioural, neuropsychological, and psychosocial measures from a cohort of 11,875 youths over the 10 years of their adolescent development.In this paper, we used the data from 6,331 of the ABCD sample.The data sample was reduced from 11,875 to 6,331 as we excluded all the participants missing key variables or with low attention/effort performance on any of the investigated measures consistent with the quality assurance performed on the HCP data and those done on similar studies in the literature using the same dataset (Adise et al., 2021).Table 2 shows the demographics of the investigated sample.Refer to section 2.3.1 on quality control for more information.Briefly, however, the initial number of participants was 11,875, but only 11,368 of them had delay discounting data.
After filtering for consistent delay discounting, 9,511 participants remained.This number was further reduced to 8,699 individuals who had available N-back fMRI scans.It was then narrowed down to 8,531 after including those who had taken the National Institute of Health (NIH) Toolbox List Sorting Task.Subsequently, the number decreased to 6,829 for participants with available parental education data and further to 6,331 for those with available parental income data.The brain scans and the NIH neurocognitive measures were collected only during the baseline and the second follow-up, while the DRD data were only collected during the first and the third follow-ups.To deal with the different time points records, the readings from each subject were averaged to constitute a single number that represents the readings within the 4 years period.Due to this drastic drop in the sample size, a sensitivity analysis was performed to assess the statistical difference between the included and excluded samples in terms of J o u r n a l P r e -p r o o f demographics (i.e., sex at birth, parental education and income) and the working memory variables (i.e., in-scanner N-back performance and the list sorting task score).

Delayed Reward Discounting
Human Connectome Project.There are several evaluation techniques for assessing DRD.The most commonly used metrics to describe the DRD is calculating the Area Under the Curve (AUC; Myerson et al., 2001) where the curve is constructed by plotting the indifference points over the delay periods and the delay discounting rate (k), which is a parameter in the hyperbolic function that describes how steeply an individual discounts commodities across time.There were two DRD tasks in HCP; the first used a larger delayed amount of $200 and the second of $40,000.Both tasks used an adaptive adjusting-amount approach in which the delay time was held constant while the immediate dollar amount varied on a trial-by-trial basis in accordance with the obtained responses from the participants (Estle et al., 2006;Green et al., 2007).The participants' responses are then used to calculate the indifference points that are used calculate both k and AUC.Additionally, due to the fact that AUC has shown to give a disproportionate contribution of indifference points at long delays to the total AUC, with minimal contribution from indifference points at short delays, we decided to add to the results a transformed AUC measure AUClogD as proposed by Borges et al. (2016).

Adolescent Brain Cognitive Development Study. A 42-item adjusting amount DRD
task developed based on the procedure proposed by Koffarnus & Bickel (2014) was implemented in the ABCD, where the child was presented with choices between a small reward given immediately versus a larger reward of $100 reward given at different points in the future (6 hours, 1 day, 1 week, 1 month, 3 months, 1 year, and 5 years).For the user manual used for the DRD procedure in ABCD J o u r n a l P r e -p r o o f (https://www.millisecond.com/download/library/v6/delaydiscountingtask/).The same methods used to calculate k, AUC, and AUClogD in the HCP dataset were used here.

Imaging Parameters
Human Connectome Project.In HCP data collection, a 3T Siemens Skyra scanner (Siemens AG, Erlanger, Germany) with a 32-channel head coil at Washington University in St. Louis was used to acquire the fMRI signals images with an isotropic voxel of the size of 2.0mm, Field of View (FOV) of 208 × 180 mm, flip angle of 52 degrees, matrix of 320×320, 256 sagittal slices, Repetition Time (T.R.) of 720 ms, Time to Echo (T.E.) of 33.1 ms, and a multi-band acceleration factor of 8. See (Glasser et al., 2013;Van Essen et al., 2013) for more details on the acquisition parameters, reconstruction, and pre-processing of the collected fMRI signals.All acquired scans were manually checked by a technician after acquisition for their quality.For more information about the HCP quality control procedure, refer to (Marcus et al., 2013).The scan was repeated twice for each participant; one was acquired from the right hemisphere to the left and the other from left to right.The acquired signals were then preprocessed by the HCP team using a minimal pre-processing pipeline (Glasser et al., 2013), which included some motion correction, gradient unwarping, field-map-based EPI distortion correction, registration to the structural scan and into MNI152 space as well as grand-mean intensity normalization.

Adolescent Brain Cognitive Development Study.
The ABCD data was collected from 21 different research sites around the United States.All the participants completed MRI and underwent a full screening at the age of 9-10 to serve as a baseline for the following years.For more information about the procedures of data collection, protocols, registration, and preprocessing, see Casey et al. (2018).The fMRI signals were pre-processed by the ABCD team using a similar minimal pre-processing pipeline to HCP.The fMRI images were then J o u r n a l P r e -p r o o f parcellated into different Regions of Interest (ROIs) using the Desikan atlas (Desikan et al., 2006a) in FreeSurfer (Fischl, 2012).For more information about pre-processing, refer to (Hagler et al., 2019).For the scanners' parameters, see: (https://abcdstudy.org/images/Protocol_Imaging_Sequences.pdf).

N-Back fMRI Task
Human Connectome Project.In the HCP N-back fMRI task, the participants were given blocks of trials that consisted of places, tools, faces, and body parts.During each run, the 4 different stimulus types were shown to the participants in separate blocks.Each two runs included 8 N-back task blocks, each of them lasting for 27.5 s.The 8 blocks included four 0back blocks, four 2-back blocks, and four resting.Each resting block lasted for 15 s each.An additional 2.5s were taken at the beginning of each block for the purpose of informing the participants which task followed (i.e., 0-back or 2-back).A total of 10 trials, each lasting 2.5s, were conducted.
Adolescent Brain Cognitive Development.In the ABCD dataset, the N-back task was slightly modified and called emotional EN-back, where elements of facial and emotional processing were incorporated into the task.Similar to the HCP N-back task, the ABCD ENback task consisted of 8 blocks, each with 10 trials and the stimuli were presented for 2.5s, followed by a plus sign that was shown for 1s.The stimulus included houses and happy, fearful, or neutral faces.

The List Sorting Working Memory Task
The NIH toolbox (Weintraub et al., 2013) was administered in both the HCP and ABCD studies, as they are validated for an age range of 3-85 years, and it includes the list sorting task.

J o u r n a l P r e -p r o o f
Similar to the N-back fMRI task, the list sorting task assesses the WM.However, it uses a different approach to this and, as such, likely measures somewhat different aspects of WM.
During the task, the participant is asked to sequence different visually and orally presented pictures (e.g., animals and foods) with a sound clip and written text of the item's name.The participant is then asked to create two lists; the first has items sequenced or arranged based on the size from smallest to largest, and the second is based on the size and type.For example, the list should start with animals arranged by size from the smallest to the largest, followed by the foods arranged based on the size from smallest to largest.

Quality Control
Delayed Reward Discounting.The DRD data was tested in both datasets, and the inconsistent responses were eliminated from the analysis.In our analysis, the participant's response was deemed inconsistent if it had more than 3 inconsistencies, and a single inconsistency was defined as having an indifference point for a given delay that was larger than that of an indifference point for a shorter delay (Owens et al., 2017;Owens et al., 2021).For example, a larger indifference point for a 1-week delay than for 1-day delay would be an inconsistency because indifference points should always decrease with the increase of the delay period (Owens et al., 2017;Owens et al., 2021).In addition to this primary quality control pipeline, two secondary pipelines were used.The first was based on how well the indifference points fit the hyperbolic regression model and the second was based on the rules proposed by Johnson and Bickel (2008), which states any DRD readings that meet one or both of following two rules should be deemed invalid: 1) if a later indifference point is larger than the earlier one by 20% of the later reward; 2) if the difference between the first and last indifference point is smaller than 10% of the later reward.In the HCP dataset, 1058 participants had their DRD data available but was reduced to 1038 participants after eliminating those who had inconsistent J o u r n a l P r e -p r o o f responses.In the ABCD dataset, the reduction was much more significant as the number of participants who had their DRD data available was 11368 which was reduced to only 9511 who had consistent DRD task responses.This significant reduction may relate to the age of participants, who were significantly younger and may have not fully understood the task.
N-back fMRI and the List Sorting Task.A number of manual and automated quality control procedures were implemented by the corresponding data collectors and releasers.For more information about the full quality control pipeline in ABCD and the HCP datasets, refer to (Hagler et al., 2019;Marcus et al., 2013), respectively.Participants without N-back fMRI data were removed, reducing the total number of participants to 8671 in the ABCD data and 1001 in the HCP dataset.Similarly, we eliminated all the participants who did not have their list sorting score available, as well as those who did not have reported data in any of the used covariates (biological sex, age, income, and education).The elimination of the subjects with unavailable covariates resulted in no reduction in the HCP, but it resulted in reducing the ABCD data subjects from 8671 to 6331.

fMRI Processing
The minimally pre-processed (Glasser et al., 2013) HCP fMRI data was downloaded, and Analysis of Functional NeuroImages software (Cox et al., 1996) was used to spatially smooth the data using a 6mm full-width half-maximum Gaussian filter.Subsequently, general linear modelling was conducted using regressors for each condition (2-back, 0-back, and instruction screens).Additionally, activation during 2-back was contrasted to a baseline of activation during the 0-back task, and the average beta weights from this contrast were derived for each subject.In the HCP analysis, regions of interest (ROIs) were based on clusters found in Owens et al. (2019), which was conducted on the same dataset.The ROIs represent regions in which there is the most pronounced activation during the 2-back (relative to 0-back).

J o u r n a l P r e -p r o o f
The ABCD fMRI data for activation during 2-back (relative to 0-back) have been fully processed and parcellated to the Desikan atlas (Desikan et al., 2006b) for the cortical regions and the ASEG atlas (Fischl et al., 2002) for the subcortical regions by (Hagler et al., 2019).
Consistent with the HCP analysis, we used the same method and parameters used in (Owens et al., 2018) to identify the ROIs that correspond to the working memory activation in the ABCD N-back task.

Mixed Effects Regression Models and Multiple Comparisons
Linear mixed-effects models were used in our analysis to account for the mixed effects, where mixed effects refer to the mixture of fixed and random effects, which is generally presented in Equation 1 (Harrison et al., 2018).
Where y is the dependent variable x is the independent variable (fixed effect), and g is a random intercept for each group.This random intercept included the participant's family and subject ID in HCP and the participant's family and scanning device for ABCD.
This approach was selected as both HCP and ABCD datasets contain twin pairs, which introduces a reduction of variability.Additionally, the ABCD data was recorded at multiple sites, which adds even more variability (Casey et al., 2018;Van Essen et al., 2013) and the HCP data had two tasks of DRD that had different reward magnitudes (i.e., $200 and $40000).
In these analyses, we used the false discovery rate correction to reduce the inflation of Type-I errors (Benjamini & Hochberg, 1995).In order to make sure that the results of our analyses are not driven by the covariates used, we examined the relationship with and without covariates, where the considered covariates were education, income, age, and biological sex.
Thes covariates were particularly considered due to their effect on either or both DRD and WM (Button et al., 2023;Johnson et al., 2021;Reimers et al., 2009;Sloan et al., 2023).Essentially, the mixed effects models were used to investigate the relationship between the ROIs activation This is done to determine the effect size caused exclusively by the investigated independent variable.
To complement our analyses with a holistic approach, a principal component analysis N-back score, list sorting score, and the DRD area under the curve) was examined using mixed effects models.Furthermore, the ROIs activations were then examined using the first principal component for each of the datasets using PCA, and this aggregated variable in relation to the behavioural measures of interest was once again investigated using mixed effects models.
Finally, before the variables were fed to the linear mixed-effects models, they were tested for normality and were winsorised and transformed if needed to improve the normality of the data.
J o u r n a l P r e -p r o o f

Human Connectome Project
In the HCP analysis, when controlling for age, sex, income, and education, we found no significant associations between the DRD matrices (i.e., log(k), AUC, and AUClogD) and the activation of brain regions induced by the 2-back task relative to the 0-back task during the fMRI scan.An association was found between the WM activations and the list sorting task score and the in-scanner N-back task score, suggesting the neural correlates were valid assays of WM.Table 3 shows the FDR corrected P-value and ΔR 2 of all the measures and ROIs as well as the chosen component from the PCA.Principal component (PC) 1 was chosen as it explains 71% of the common variance of the ROIs, and it had an eigenvalue of 2.66.The same findings were replicated using the DRD raw data, the filtered DRD data using the hyperbolic model fit, and Johnson and Bickel (2008) quality control algorithm, see Supplemental Tables S2, S4 and S6.For more results, such as the unadjusted P-values and the beta weights of the primary analysis or the PCA rotations, refer to Supplemental Tables S8-S17.
Similarly, when no covariates were included, the results were still consistent with the findings of analysis done with covariates.The correlation of all the behavioural measures and the PCA aggregated component were investigated and as expected both in-scanner and out of scanner working memory measures were correlated with the aggregated component that represents the variance of ROIs activations while the DRD matrices (i.e., log(k), AUC, and AUClogD) were not, see Figure 1.Interestingly, both list sorting task score and the in-scanner N-back task score were found to be correlated with the DRD matrices.

Adolescent Brain Cognitive Development Study
The sensitivity analysis between the included and excluded samples showed a significant difference between the investigated variables and demographics (i.e., sex at birth, J o u r n a l P r e -p r o o f parental education, income, in-scanner N-back performance, and the list sorting task score), see Table S1 for more details.Consistent with the analyses conducted on the HCP data, the ABCD study results showed no association between the WM activations and DRD scores.
Additionally, the associations of the list sorting task score and the in-scanner N-back task score with the N-back ROIs activations were consistent with what was found in the HCP analysis, with some differences in the effect sizes.Table 5 shows the corrected P-value and the ΔR 2 for the investigated variables while considering the covariates.For more results on the ABCD analysis, such as the results without considering the covariates, beta weights, uncorrected Pvalues, and the PCA rotations, refer to Supplemental Tables S18-S26.Together with the empirical ROIs of ABCD, Table 5 shows the relationship between PC1 and the other investigated behavioural measures.Where PC1 explains 60% of the common variance of the ROIs, and it has an eigenvalue of 3.28.The findings were consistent when analysed with the raw unfiltered and filtered DRD data, employing both the hyperbolic model fit and the Johnson and Bickel (2008) quality control algorithm.For detailed results, refer to Supplemental Tables S3, S5 and S7.The correlations of all behavioural measures with the PCA aggregated component were investigated.As expected, both in-scanner and out-of-scanner working memory measures showed a correlation with the aggregated component, representing the variance of ROIs activations.However, the DRD matrices (i.e., log(k), AUC, and AUClogD) did not exhibit this correlation.For details, see Figure 1.Similar to the HCP findings, both list sorting task score and the in-scanner N-back task score were found to be correlated with the DRD matrices, though the correlations were considerably weaker than those found in the HCP data.
J o u r n a l P r e -p r o o f

Discussion
Although the association between the neural correlates of WM and DRD is intuitive, as both are facets of executive function and have been strongly suggested by oblique metaanalytic approaches (Wesley & Bickel, 2014), our analyses did not find significant associations between brain activation during WM and DRD.These results were consistent in both large datasets and although the size of the association was affected by the number of covariates controlled for, the results were very similar with or without considering the covariates.The results were also consistent when using the raw data in the analyses and filtered data using different well established quality control pipelines.In contrast, in both datasets, the neural correlates of WM were robustly correlated with the behavioural outcomes on the in-scanner task and an independent WM task was obtained separately.These results support the validity of the WM neural correlates obtained.Finally, aggregating the ROIs using PCA produced a latent aggregate indicator of brain activity that accounted for the majority of the variance but was likewise not associated with DRD in both cases.Collectively, these results do not implicate the BOLD-signal correlates of WM directly in relation to DRD behavioural performance.
However, the correlation between the behavioural measures of WM (i.e., in-scanner N-back performance and the list sorting task score) with the DRD matrices (i.e., log(k), AUC, and AUClogD) suggests that the interaction between these two aspects (i.e., WM and DRD) of cognitive functioning is far more complex than generally assumed.
It was observed that the associations, in general, were much weaker in the ABCD dataset.This may be due to "noisier" ABCD data (e.g., collected from multiple sites and using different scanners), or maybe because the ABCD data was collected from an early adolescent population that is still undergoing substantial brain development, so there is more variability in functional activation.The association between WM brain-induced activations, the list sorting J o u r n a l P r e -p r o o f task, and the in-scanner N-back score is intuitive, expectable, and an internal validation to our overall analyses that indirectly supported the validity of our null finding.
One possibility for the divergence of the current findings from previous reports is publication bias, namely, that small sample studies reporting significant results are more likely to be published than those reporting non-significant results (Kepes et al., 2014).The intuitive overlap of these two cognitive processes may have effectively promoted underpowered but supportive studies and made it less likely for non-significant reports to be published.Small sample size poses a substantial limitation in the scientific field as it limits the ability to replicate and use the induction principle to come up with meaningful generalized findings (Poldrack et al., 2017).However, in this paper, two of the largest fMRI-collected datasets known to date were used, which also represent distinct populations from diverse backgrounds.In all the current analyses, a statistically conservative approach was taken to enhance confidence in the findings.As such, we tried to balance efforts to balance type I and type II errors by examining the association with and without considering covariates, using an effective yet sensitive multiple comparison correction (i.e., false discovery rate; Benjamini & Hochberg, 1995).
The absence of a statistically significant relationship between DRD and WM-related brain activities was supported by Moro and colleagues (2023) who suggested apparent independence between WM and DRD.This notion was supported by other publications as well (e.g., Garzón et al., 2022;Acuff, 2012).This has important implications.For instance, it may raise doubts about the effectiveness of WM training for mitigating impulsive traits.However, it is crucial to emphasize that the lack of correlation in this study does not definitively rule out any mechanistic connection between DRD and WM.It is possible that the N-back task may not fully assess the aspects of WM required to capture its association with DRD.This idea arises from the conceptual appeal and intuition of such a connection as well as the correlation between their behavioural variables.

Conclusion
In conclusion, contrary to what has been suggested in the literature and to our predictions, minimal association was found between behavioural performance and brain activations induced by the N-back fMRI task and the delay discounting indicators in both large datasets.These findings were consistent when the analyses were repeated with and without considering covariates and even when investigating different populations (i.e., adolescents and young adults).This suggests that the relationship between these two aspects (i.e., WM and DRD) of higher cognitive functioning is mechanistically far more complex than currently believed.
J o u r n a l P r e -p r o o f patterns and the Adolescent Brain Cognitive Development (ABCD) study, which is a longitudinal neuroimaging study collecting neuroimaging and cognitive data from children and adolescents to understand brain development and its relation to behaviour (Saragosa-Harris et J o u r n a l P r e -p r o o f

J
o u r n a l P r e -p r o o f induced by the N-back fMRI task and three other variables: the N-back task in-scanner score, List sorting score, and the DRD area under the curve.The measure ΔR 2 was introduced, which we calculated by subtracting the R 2 of a mixed effects model that includes the independent variable and covariates and the R 2 of a mixed effects model that included only the covariates.

(
PCA) was used to reduce all the WM ROIs activation into one component that represents the ROIs variance in each of the datasets.This single component was then used as an independent variable in the mixed effects regression model to investigate the relationship between the chosen PCA component and the other investigated dependent variables (i.e., DRD, in-scanner N-back task score, and list sorting score).Specifically, the empirical ROIs corresponding to WM activations during the N-back fMRI task have been identified in both datasets (i.e., HCP and ABCD) through the investigation of significant correlations between the in-scanner Nback score and the brain region activations.The relationship between the activations in the identified ROIs in each dataset and each of the behavioural measures of interest(i.e., in-scanner

J
o u r n a l P r e -p r o o fOne notable limitation of this study is the low quality of the DRD data in the ABCD dataset, potentially due to some participants being too young to fully understand or accurately follow the instructions of the administered DRD task.Additionally, the comparisons between the two datasets were constrained due to substantial heterogeneity in terms of age and reward values.Another limitation pertains to the absence of DRD neural activity data within both of the investigated datasets.While we extensively examined the association between DRD scores and the WM N-back neural activity data, the lack of direct neural correlates for DRD within our analysed datasets limits our ability to elucidate the underlying neural mechanisms governing this relationship.Incorporating neuroimaging measures related to DRD could offer valuable insights into the neural substrates involved in impulsive decision-making and how they interact with working memory processes.Future research endeavours should aim to address this limitation by incorporating a DRD fMRI task to provide a more comprehensive understanding of the neural underpinnings of DRD in diverse populations.
Note.Aggregated ROIs represent the first component derived from a principal component analysis.The numbers and colour coding indicate the values of the correlation coefficients.Significance levels are denoted by asterisks: * indicates p < 0.05 and ** indicates p < 0.005.

Table 3 : Mixed Effects Models Covariate-adjusted Associations between Working Memory- associated Brain Activity in relation to Working Memory or Delay Discounting in the Human Connectome Project. ROI In-Scanner Score
J o u r n a l P r e -p r o o f