The computational and neural substrates of individual differences in impulsivity under loss framework

Abstract

Numerous neuroimaging studies have identified significant individual variability in intertemporal choice, often attributed to three neural mechanisms: (1) increased reward circuit activity, (2) decreased cognitive control, and (3) prospection ability. These mechanisms that explain impulsivity, however, have been primarily studied in the gain domain. This study extends this investigation to the loss domain. We employed a hierarchical Bayesian drift‐diffusion model (DDM) and the inter‐subject representational similarity approach (IS‐RSA) to investigate the potential computational neural substrates underlying impulsivity in loss domain across two experiments (n = 155). These experiments utilized a revised intertemporal task that independently manipulated the amounts of immediate and delayed‐loss options. Behavioral results demonstrated positive correlations between the drift rate, measured by the DDM, and the impulsivity index K in Exp. 1 (n = 97) and were replicated in Exp. 2 (n = 58). Imaging analyses further revealed that the drift rate significantly mediated the relations between brain properties (e.g., prefrontal cortex activations and gray matter volume in the orbitofrontal cortex and precuneus) and K in Exp. 1. IS‐RSA analyses indicated that variability in the drift rate also mediated the associations between inter‐subject variations in activation patterns and individual differences in K. These findings suggest that individuals with similar impulsivity levels are likely to exhibit similar value processing patterns, providing a potential explanation for individual differences in impulsivity within a loss framework.

The current study employed a hierarchical Bayesian drift‐diffusion model and inter‐subject representational similarity approaches across two experiments to investigate the neural substrates underlying intertemporal choices under loss framework. The findings underscore the multifaceted nature of impulsivity, highlighting the interplay between computational and neural factors.

1. INTRODUCTION

Intertemporal choice refers to the process of deciding between either an immediate, smaller reward or a delayed, larger reward (Ainslie, 1975) and plays a pivotal role in various aspects of life, including academic achievement, success, and coping with frustration and stress (Mischel et al., 1989). The ability to make such choices, often assessed by a discounting rate K, has been closely associated with a range of variables. These include sociodemographic factors (Eppinger et al., 2017), personality traits (Yeh et al., 2021), cognitive styles (Peters & Büchel, 2011), inhibitory control (Figner et al., 2010), and intelligence (Shamosh & Gray, 2008). Notably, individuals with a larger K are not only more likely to make impulsive decisions (Wang et al., 2016), but also exhibit behavioral tendencies symptomatic of psychiatric disorders, such as gambling (Alessi & Petry, 2003), drug abuse (Bickel & Marsch, 2001), smoking (Bickel et al., 1999), and Parkinson's disease (Antonelli et al., 2014).

Variations in brain structure, activity, and connectivity patterns have been proposed to explain individual differences in intertemporal choice, particularly within the gain domain (Peters & Büchel, 2011). The reward circuit, cognitive control, and prospection network comprise the three key neural networks in an integrated framework for understanding impulsivity (Peters & Büchel, 2011). Specifically, increased brain activations and changes in gray matter volume (GMV) in the reward circuit, encompassing the ventral striatum, ventromedial prefrontal cortex, and orbitofrontal cortex (OFC), have been linked to impulsive decision‐making (Ballard & Knutson, 2009; McClure et al., 2004; Wang et al., 2016). Furthermore, a decrease in top‐down cognitive control (e.g., mediated by the prefrontal‐parietal cortex) and prospection (e.g., hub regions, such as the hippocampus) can account for why some individuals exhibit greater impulsivity when faced with temptation (Figner et al., 2010; Peters & Büchel, 2009). Recent studies have suggested that attribute‐related neural activity (e.g., reward amount, delay time length, and discounted probability) and corresponding network properties (e.g., local efficiency) can also predict individual impulsive performance (Wang, Wang, et al., 2021; Wang, Zhang, et al., 2023). In addition, previous studies have demonstrated specific neural activity in the three networks mentioned above, which are critical for underlying the neural mechanisms of impulsivity. This includes findings from electroencephalography (EEG) (Lin et al., 2018), intracranial EEG (Gui et al., 2018), and functional near‐infrared spectroscopy (Heinzel et al., 2013).

Despite these advances, however, much of the literature has been centered around the discounting of future gains, limiting the extent to which the same theoretical framework can be applied to the discounting of future losses. Indeed, several studies have documented distinct cognitive/emotional processes and neural substrates underlying the discounting of future losses, such as affect (Harris, 2012; Xu et al., 2009), attribute‐processing‐related asymmetries (Tanaka et al., 2014), and distinct valuation systems (Zhang et al., 2018). Recent research has identified two distinct groups of choice preferences and has shed light on the potential cognitive and neural mechanisms. These mechanisms include a changed valuation for delayed‐loss options, atypical action selection processing, and a decrease in functional coupling between these two processes (Wang, Zhang, et al., 2023).

Beyond distinct neural underpinning within the gain and loss frameworks, various behavioral and psychological manifestations emerge, including the magnitude effect, feelings of dread, and gain‐loss asymmetry, particularly in psychiatric disorders such as nicotine dependence. Previous studies have shown that delayed gains are discounted significantly more steeply than delayed losses (Estle et al., 2006). Furthermore, individuals tend to experience a feeling of dread when facing future losses, which may not apply to future gains (Harris, 2012). Notably, individuals with nicotine dependence exhibit a significant association with delayed gains but not losses (Mejía‐Cruz et al., 2016). Although much of the existing literature has focused on comparing psychological processes and neural substrates within the gain and loss framework, the extent of individual differences in impulsivity under the loss framework remains largely unexplored. Investigating individual variability in impulsivity within the context of the loss framework can help fill this research gap and provide a more comprehensive understanding of inter‐temporal choices.

We tackled the investigation of individual variability in impulsivity using the innovative analytical approach, inter‐subject representational similarity approach (IS‐RSA). IS‐RSA examines whether individuals with similar behavioral responses, personality traits, or psychological processes exhibit neural similarity in functional activations, morphological characteristics, and connectivity (Finn et al., 2020; Thirion et al., 2006). This approach emphasizes the geometric properties in a high‐dimensional space constructed by brain or behavioral patterns and employs a second‐order isomorphism to establish associations across individuals (Kriegeskorte & Kievit, 2013). Second‐order isomorphism suggests that there should still be an approximate parallelism between different internal representations and their corresponding external objects (Gordon & Hayward, 1973). In cognitive science and neuroscience, there is a higher level of similarity or correspondence between different cognitive processes or neural structures. This isomorphism may not be limited to formal similarities, but also includes similarities in function, dynamic changes, or information processing mechanisms. In other words, IS‐RSA is sensitive to phenotypic differences between individuals and offers greater detection of subtle differences. IS‐RSA has been utilized in various research domains, including moral decision‐making (van Baar et al., 2019), emotion (Chen et al., 2020), narratives (Nguyen et al., 2019), and personality (Liu et al., 2019). Thus, we applied IS‐RSA to investigate individual variability in impulsivity for a more nuanced, comprehensive, and sensitive approach than traditional methods.

Considering the significant research bias between gain and loss domains in intertemporal choices, particularly regarding loss‐related individual differences, we aim to investigate its computational and neural substrates from an exploratory perspective. We, therefore, approached this investigation using a combination of the drift‐diffusion model (DDM) and the IS‐RSA approach, applied to a revised intertemporal choice task under the loss framework. In this task, the amounts of the immediate and delayed loss options were independently manipulated, and the delay time length for future options was fixed (e.g., half a year later). The DDM model consists of four parameters (i.e., drift rate v, decision boundary a, nondecision time T, starting point z) and can be used to uncover the latent cognitive processing of decision‐making assessed by the evidence accumulation process (Ratcliff et al., 2016). The use of a computational model approach can enhance the ability to predict individual differences in impulsivity by allowing for the quantification of latent cognitive processes, such as the evidence accumulation process assessed by the drift rate in the DDM, thereby improving the precision and reliability of predictive models.

Taken together, we hypothesize the following:

1. Impulsive individuals are expected to exhibit a larger drift rate v in loss situation. This hypothesis builds upon a previous study that predicted differences in drift rate, particularly related to amount and time, correlating with impulsivity within a gain framework (Amasino et al., 2019). We hypothesized that, within a loss‐oriented framework, impulsive individuals tend to accumulate evidence more rapidly to facilitate quick decision‐making, thereby demonstrating elevated drift rates.

2. Brain activations in response to loss amounts and GMV are predicted to correlate with individual K, as measured via traditional linear regression analysis. Previous research consistently identified robust associations between amount‐related brain activations and K in gain domain (Wang, Chen, et al., 2023; Wang, Wang, et al., 2021). Extending this hypothesis to the realm of losses, we proposed that brain activation pattern triggered by varying loss amounts might reflect individual variations in impulsivity. This relationship is particularly pertinent to impulsivity‐associated neural structures, such as GMV.

3. Individuals with similar impulsivity are anticipated to exhibit similar neural activation patterns, especially in response to processing varying amounts of loss. This phenomenon will be quantitatively assessed using IS‐RSA. Specifically, our hypothesis posits that individuals who share similar levels of impulsivity will display consistency in the brain regions activated and the intensity of activity in these regions when confronted with differing levels of loss stimuli.

4. The latent cognitive process assessed by the drift rate is hypothesized to mediate the association between individual differences in brain properties and variability in impulsivity. This hypothesis is based on the findings suggesting that drift rate can offer insights into decision processing, especially within a gain framework. This coherence significantly enhances our understanding of how impulsivity influences neural processing. Furthermore, we hypothesized that drift rate serves as a crucial intermediary linking brain attributes, such as GMV and functional connectivity, to impulsivity traits within a loss framework. This approach promises a deeper understanding of the mechanisms through which impulsivity shapes intertemporal choice behavior.

2. METHODS

2.1. Participants

One hundred and five participants (42.9% males; ages 17 to 25 years old) took part in the present study (Exp. 1). Eight participants were excluded from further analyses due to misunderstanding the task (n = 3) or large head motion (frame‐wise displacement [FD] >0.5 mm in any one of three runs; n = 5), for a final sample of 97 (FD; M ± SD = 0.103 ± 0.045 mm; range = [0.057–0.356 mm]) in the analysis. Another 58 participants (17 males; ages 18–23 years old) without brain imaging data were additionally recruited to replicate behavioral findings from Exp. 1 (Exp. 2). Only participants without any neurological or psychiatric history were examined in this study. We obtained written informed consent from all adult participants (age 18–25) before formal investigation. For adolescent participants (age < 18), consent forms were signed after receiving verbal consent from their parents. The current study was approved by the institutional review board of the Faculty of Psychology at Tianjin Normal University, China.

2.2. Intertemporal choice task in loss framework

In the interest of brevity, we report only the key details of the Intertemporal Choice Task below. More comprehensive details about the task procedures can be found in Wang, Zhang, et al. (2023). When individuals defer their decision‐making to a future point in time, such as 6 months later, they are more likely to assess the problem from a holistic perspective. By deliberately selecting a delayed loss option and rationalizing the timing of this decision, individuals can leverage the benefits of psychological distance, ultimately enhancing the quality and effectiveness of their choices (Trope & Liberman, 2003). Participants were asked to choose between an immediate‐loss option (16 levels: 25–100) and a delayed‐loss option (fixed at half a year later; 16‐levels: 28–112). Such time manipulations were designed to facilitate the computation of subjective value for participants. The task was divided into three runs, comprising 256 combinations. The wide range of options was designed to elicit a precise preference, ranging from strong acceptance to strong rejection of immediate loss over delayed loss. This design aimed to accurately capture the distributed representation of the brain and identify linear activation patterns associated with different values. We used Optseq2 to determine the timing and order of stimulus presentation in the event‐related design, optimizing for estimation efficiency (Dale, 1999). Upon completion of the experiment, participants received final compensation based on a predetermined rule to enhance ecological validity (see (Wang, Zhang, et al., 2023)).

2.3. Behavioral data analysis

We used logistic regression to fit the behavioral data based on prior studies (Wang et al., 2014; Wang, Chen, et al., 2023; Wang, Zhang, et al., 2023), where the independent variables were the magnitudes of the immediate and delayed‐loss, and the dependent variable was the choice of immediate and delayed‐loss option (coded as 0 or 1). The impulsivity index (K) was indirectly computed as the absolute ratio between the weight of the delayed‐loss option and the weight of the immediate‐loss option (| βDL/βIM |). A larger K value indicates a stronger preference for the immediate‐loss option, representing a form of impulsivity performance within the loss framework. Here, it should be noted that K is not equal to discounting rate estimated by discounting model (e.g., hyperbolic function).

2.4. DDM analysis

The DDM is a key computational model used to assess two‐choice decision‐making processes operating via four critical parameters: (1) the decision threshold (a), (2) drift rate (v), (3) starting point (z), and (4) nondecision time (T) (Figure 1c). This model characterizes decisions as noisy evidence accumulation processes that fit both reaction time (RT) and accuracy. The drift rate represents the speed of evidence accumulation toward a decision boundary. To obtain individual parameters under the loss framework, we used a hierarchical Bayesian DDM model to fit individual decisions and response times. This approach allowed us to explore potential decision‐dynamic processes for impulsive individuals and further distinguish the possible influences of impulsivity on distinct decision‐dynamic processes.

FIGURE 1.

Behavioral results. Color‐coded loss options (a) and response time (b) at each level of immediate/delayed‐loss‐amount combination were displayed in two experiments (Exp.1: n = 97; Exp.2: n = 58). (c) An illustration of the latent decision dynamic process for DDM with three parameters, such as drift rate (v), decision boundary (a), and nondecision time (T). In this model, a Gaussian noisy evidence accumulation process will terminate when the accumulated evidence reaches one of two response boundaries (e.g., immediate or delayed loss options). (d) The values of fitting parameters in the full model were plotted in both experiments. The drift rate was positively correlated with impulsivity index K (e) and the proportion of choosing the immediate‐loss option (f). P_IM: proportion of choosing immediate‐loss option; IML, immediate‐loss amount; DLL, delayed‐loss amount.

The Bayesian estimation generates the joint posterior distribution of the full model parameters based on observed data providing not only a point estimate but also includes the uncertainty of the estimate to enable Bayesian inference (Gelman, 2005). We further utilized Markov Chain Monte Carlo simulations to generate 120,000 samples from the joint posterior parameter distributions (Dani Gamerman, 2006). The first 50,000 samples were discarded as burn‐in (Matzke et al., 2013), and a thinning factor of 10 was applied, with outliers specified at 5%. Convergence was estimated by visually inspecting the Markov chains and computing the R‐hat Gelman‐Rubin statistic (v: 0.9999; a: 1.0001; t: 1.0000; z: 1.0000), with successful convergence determined based on values less than 1.05 (Krypotos et al., 2015). After comparing the deviance information criterion (DIC) of each model, the best model can be determined by lower DIC values (Spiegelhalter et al., 2002). In this model, we calculated four model‐related indices to perform subsequent analyses, including decision threshold, drift rate, starting point, and nondecision time.

2.5. fMRI data collection and preprocessing

2.5.1. Brain imaging

All brain‐imaging scans were acquired using a Siemens Prisma 3.0 T scanner equipped with a 64‐channel head coil at the Center for MRI Research at Tianjin Normal University. Subjects laid supine on the scanner bed and viewed visual stimuli back‐projected onto a screen through a mirror attached to the head coil. The experiment was programmed and run with MATLAB (MathWorks) and Psychtoolbox‐3 with version of 3.1 on a PC laptop. Subjects' responses were collected using an MRI‐compatible button box.

For each functional session, T2*‐weighted functional images were acquired with a simultaneous multi‐slice sequence. The specific parameters used were the following: repetition time (TR) = 2000 ms; echo time (TE) = 30 ms; GRAPPA factor = 2; multiband acceleration factor = 2; flip angle = 90°; field‐of‐view (FOV) = 244 × 244 mm²; slice thickness = 2 mm; slice gap = 0.3 mm; voxel size = 2 × 2 × 2 mm³. The scan time of each run was 8 min, and thus 242 volumes were acquired for each run. The slices were tilted approximately 30°clockwise from the AC‐PC plane to obtain better signals in the OFC. Additionally, MPRAGE T1‐weighted image for each subject was also obtained (192 slices, TR = 2.530 ms, TE = 2.98 ms, multiband factor = 2, flip angle = 7°, FOV = 224 × 256 mm²; voxel size = 0.5 × 0.5 × 1 mm³).

2.5.2. Imaging preprocessing

The fMRI Expert Analysis Tool (version 5.98, part of the FSL package; http://www.fmrib.ox.ac.uk/fsl) was employed for imaging preprocessing and statistical analyses. Standard preprocessing steps, including slice‐timing, motion correction, filtering, registration, and smoothing, were performed following past research (Deng et al., 2023; Wang et al., 2014; Wang, Chen, et al., 2023). In the generalized linear model (GLM), four parametric regressors were included at the first level: (1) the overall task regressor (1 for each trial); (2) the magnitude of the immediate reward loss; (3) the magnitude of the delayed reward loss; (4) RT. All regressors, except for the task regressor, were standardized to the same range (−1 vs. 1) and then convolved with the double‐gamma canonical hemodynamic response function. These regressors were estimated during the decision‐making period, which commenced with the presentation of stimulus alternatives and concluded when the participant responded. Trials without a valid response and six motion parameters for head movement were further modeled as regressors of no interest. The second‐level analysis was conducted using a fixed‐effect model, combining all three functional runs within each participant. These second‐level results were subsequently entered into a random‐effects model for group analysis and regression analysis for each participant's impulsivity index K using a FLAME1 model. To further validate the potential impact of the original K distribution, we also conducted a second‐level analysis using the log‐transformed K. The main findings were consistently replicated (Figure S2) but were not included in the main text. Given the influence of head motion on functional activations, the FD was additionally incorporated as a confounding factor in the main GLM. Group images were thresholded using cluster detection statistics, with a height threshold of z > 3.1 and a cluster probability of p < .05, corrected for whole‐brain multiple comparisons using Gaussian random field theory.

2.5.3. Percentage signal change evaluation

To assess signal change for each level of immediate loss and delayed‐loss amount, two additional models were constructed to estimate the brain signal change for each of the 16 levels of immediate and delayed reward loss, respectively. In each model, all trials with the same immediate (or delayed) losses were grouped as separate regressors (16 in total), and the delayed‐loss (or immediate‐loss) value and the RT were included as covariates of no interest. The smoothed data were utilized to fit the value function. For each ROI, we extracted parameter estimates (betas) of each level of immediate loss and delayed‐loss amount from the fitting model, averaging across all voxels in each ROI for each subject. Subsequently, percentage signal changes were computed using the formula: [contrast image/(mean of run)] × ppheight × 100%, where ppheight represents the peak height of the hemodynamic response compared to the baseline level of activity.

2.6. Brain image analysis

2.6.1. IS‐RSA analysis

The IS‐RSA was employed to examine the correlation between individual variability in brain activations and behavioral performance. First, we calculated the pattern dissimilarity of behavioral performance on impulsivity by determining the absolute differences between each pair of participants on impulsivity K or drift rate (Finn et al., 2020). Second, we computed the neural pattern dissimilarity using the representational dissimilarity matrix for each of the 200 brain parcels derived from a whole brain parcellation based on meta‐analytic functional coactivation of the neurosynth database (http://neurovault.org/images/39711). Using such a parcellation scheme offers several advantages over the conventional searchlight approach, including lower computational costs and better reflection of the properties of functional neuroanatomy. Pearson's correlation was used to perform pairwise correlation dissimilarity between each pair of participants (i.e., 1 − r), with a particular focus on un‐smooth participants' beta maps separated for immediate‐loss and delayed‐loss conditions. We then calculated the correlation between each parcel dissimilarity matrix and the behavioral dissimilarity matrix using Pearson's correlation on the lower triangle of the matrices. In addition, Spearman's correlation was further used to validate the robustness of correlation analysis, and we observed similar findings (Table S1). For significance tests, we employed two strategies to obtain the corresponding statistical values: (1) a Bonferroni‐based correction where the corresponding p‐values were multiplied by the number of parcels (200) and (2) a permutation test where we shuffled the order of the behavioral matrix 1000 times and calculated the permuted ρ value corresponding to p = .05. The statistical tests are considered valid if the actual ρ value exceeds the permuted ρ value. To further mitigate the potential impacts of confounding variables such as age, parents' educational levels, we reconducted the aforementioned analyses after controlling for these variables. The findings were replicated once more.

2.6.2. VBM analysis

T1 image data were analyzed using the voxel‐based morphometry toolbox implemented in FSL (FSL‐VBM). Additional preprocessing details, which primarily include tissue‐type segmentation, registration, Jacobian correction, and smoothing with an isotropic Gaussian kernel with a 3 mm standard deviation, can be found in Wang, Wei, et al. (2021) and Zhu et al. (2023). We used a mixed‐effect FLAME1 model implemented in FSL at the whole‐brain level to examine group differences after controlling for parental education, age at MRI scan, sex, and total intracranial volume (TIV). Statistical results were determined at the cluster level (z > 3.1, p < .001) with a family‐wise error rate of 0.05 for the correction of multiple comparisons using Gaussian random field theory.

2.6.3. Mediation analysis

We conducted several mediation analyses to investigate whether the drift rate from the computational model mediated the associations between brain morphological/functional features and impulsivity K. Linear regression analysis tested the association between (1) the brain index (e.g., GMVs in the OFC, precuneus, and cerebellum; brain activations in the dorsomedial prefrontal cortex [DMPFC], DLPFC respectively processing immediate and delayed‐loss) and K (Y = d₁ + cX + ε₁); (2) the brain index and drift rate (M = d₂ + aX + ε₂); and (3) the brain index and K with a mediator (Y = d₃ + bX + b_M + ε₃). In these equations, Y represents the outcome variable, X represents the explanatory variable, and M represents the mediator. The indirect effect was estimated as a × b. Bootstrap simulation (n = 1000) was performed using PROCESS v2.16.3 implemented in SPSS (Version 25.0) (Hayes, 2013) to test the significance of the mediation effect. Additionally, the analyses above were further conducted after controlling for confounding variables such as age, parental education levels, and gender.

3. RESULTS

In Exp. 1, individual K ranged from 0.376 to 1.385. The mean modeled accuracy of prediction was significantly higher than chance (M ± SD = 92.21% ± 5.22%; t ₍₉₆₎ = 79.66, p < .001). DDM model was used to fit individual decision preference (Figure 1c) to yield four parameters (Figure 1d), including the decision threshold (a; M ± SD = 1.618 ± 0.209), starting point (z; M ± SD = 0.534 ± 0.026), drift rate (v; M ± SD = −0.049 ± 0.312), and nondecision time (T; M ± SD = 0.938 ± 0.178). Individual K was positively correlated with drift rate (r = .967, p < .001, Figure 1e), even after controlling for confounding variables (i.e., age, gender, parental education; r = .965, p < .001), but not for the other fitting parameters (p values >.063). We replicated the analyses in Exp. 2 (see Figure 1d for the model's fitting parameters) and found a consistent correlation between individual K (range: 0.868–1.136) and drift rate (range: 0.868–1.136), r = .547, p < .001, Figure 1e. Further, the drift rate was positively correlated with the proportion of choosing the immediate‐loss option in both Exp. 1 (r = .966, p < .001, Figure 1f) and Exp. 2 (r = .918, p < .001, Figure 1f). However, we did not observe any gender differences on K in either Exp. 1 (t ₍₉₅₎ = 1.237, p = .219) or Exp. 2 (t ₍₅₆₎ = 1.298, p = .200). To reduce the impact of original K distribution, we further conducted log‐transformed K and found similar results (see Supplementary Material Results for details).

Figure 1 provides the average RTs and proportions of choosing the immediate‐loss options across a 4 × 4 matrix in both Exp. 1 and 2. The proportion of choosing the immediate‐loss option decreased with the amount of immediate‐loss but increased with the size of the delayed‐loss options (Figure 1a). Participants also exhibited faster responses when the relative differences between the two options increased (Figure 1b). Moreover, similar behavioral patterns were also replicated in Exp. 2 (Figure 1a,b).

3.1. Brain activations responding to immediate and delayed‐loss amount

We first investigated the neural substrates underlying the immediate and delayed‐loss amounts using traditional univariate analysis in Exp. 1. For the amount of the immediate‐loss, the whole‐brain analyses revealed it was positively associated with brain activations in the visual pathway, including the left occipital pole (OP; peak MNI: x = −10, y = −92, z = −6, Z = 7.11, cluster size = 2048) and the right occipital fusiform gyrus (xyz = 22, −88, −8, Z = 5.46, cluster size = 606), but was negatively associated with brain activations in the precuneus (xyz = 0, −54, 70, Z = 5.47, cluster size = 1050), left superior parietal lobule, left superior parietal lobule (SPL; xyz = −42, −46, 64, Z = 5.01, cluster size = 466), right SPL (xyz = 42, −44, 50, Z = 4.42, cluster size = 306), left OFC (xyz = −32, 30, 0, Z = 5.22, cluster size = 261), right cerebellum (xyz = 44, −56, −42, Z = 5.10, cluster size = 177), left frontal pole (FP; xyz = −20, 60, −6, Z = 4.54, cluster size = 166), right frontal operculum cortex (xyz = 46, 18, 2, Z = 3.99, cluster size = 145), and left cerebellum (xyz = −26, −70, −26, Z = 4.54, cluster size = 122) (Figure 2a; Table 1). The amount of the delayed loss was negatively correlated with brain activations in the left OP (xyz = −28, −98, 0, Z = 5.47, cluster size = 3034) and right cerebellum (xyz = 2, −56, −8, Z = 4.77, cluster size = 103) (Figure 2b; Table 1). Focusing on these regions, ROI analyses were used to validate the specific correlation between brain activation and amounts of loss and ensure the reliability of our analysis (Figure 2c).

FIGURE 2.

The brain regions modulated by the amount of the immediate and delayed‐loss options. The prefrontal‐parietal cortex and visual pathway showed sensitivity to the amount of the immediate‐loss options (a). Only the visual pathway exhibited negative correlations of its activation with the amount of delayed‐loss options (b). Percentage signal changes were linearly associated with the amount of the immediate and delayed‐loss options in several ROIs (c). The 16‐level amounts of immediate and delayed‐loss options further collapsed into eight levels to reduce noise.

TABLE 1.

Brain regions processing the amounts of immediate and delayed‐loss options.

Effect	Brain region	Cluster size (voxels)	MNI coordinates			Z
			X	Y	Z
Immediate amount	L OP	2048	‐10	−92	−6	7.11
	R OFG	606	22	−88	−8	5.46
	L SPL	466	−42	−46	64	5.01
	R SPL	306	42	−44	50	4.42
	Precuneus	1050	0	−54	70	5.47
	L OFC	261	−32	30	0	5.22
	L FP	166	−20	60	−6	4.54
	R FOC	145	46	18	2	3.99
	L cerebellum	122	−26	−70	−26	4.54
	R cerebellum	177	44	−56	−42	5.10
Delayed amount	L OP	3034	−28	−98	0	5.47
	R cerebellum	103	2	−56	−8	4.77

Note: Red and blue colors represent positive and negative correlations between brain activations and immediate/delayed‐loss amounts, respectively.

Abbreviations: FOC, frontal operculum cortex; FP, frontal pole; OFC, orbitofrontal cortex; OFG, occipital fusiform gyrus; OP, occipital pole; SPL, superior parietal lobule.

3.2. Functional substrates underlying the individual differences in impulsivity

Based on the brain activations responding to immediate and delayed‐loss amounts, we further explored whether the corresponding brain activity predicted individual variability in impulsivity. Individual K was positively correlated with the immediate‐loss‐related brain activity in the prefrontal cortex (Figure 3a; Table 2), including the left dorsomedial prefrontal cortex (DMPFC; xyz = −4, 18, 46, Z = 4.74, cluster size = 604) and right dorsal prefrontal cortex (DLPFC; xyz = 52, 28, 34, Z = 4.06, cluster size = 96). On the other hand, it was negatively associated with immediate‐loss‐related brain activity in the visual pathway (Figure 3a; Table 2), including the right OP (xyz = 14, −98, 8, Z = 4.74, cluster size = 604), left OP (xyz = −14, −96, −2, Z = 4.53, cluster size = 215), left temporal fusiform cortex (TFC; xyz = −28, −40, −14, Z = 4.74, cluster size = 184), right middle temporal gyrus (TMG; xyz = 58, −56, 6, Z = 4.43, cluster size = 160), and left parietal operculum cortex (xyz = −56, −30, 22, Z = 4.92, cluster size = 159).

FIGURE 3.

Loss‐amount‐related brain regions showing significant correlations with K and corresponding mediation effects. Immediate‐loss‐related brain activations in the DMPFC and DLPFC were positively correlated with K and their links were further mediated by the drift rate (a). Delayed‐loss‐related brain activations in the DMPFC and DLPFC were negatively correlated with K and their links were also mediated by the drift rate (b). The conjunction analysis demonstrated that drift rate mediated the associations between activation differences between immediate and delayed‐loss options and K (c). a, coefficient measuring the associations between brain activations and drift rate; b, coefficient measuring the relationships between drift rate and K; c, total effect representing the influence of brain activations on K; c′, direct effect representing the influence of brain activations on K.

TABLE 2.

Brain regions whose activations were correlated with impulsivity K.

Effect	Brain region	Cluster size (voxels)	MNI coordinates			Z
			X	Y	Z
Immediate amount vs. K	L DMPFC	376	−4	18	46	5.57
	R DLPFC	96	52	28	34	4.06
	R OP	604	14	−98	8	4.74
	L OP	215	−14	−96	−2	4.53
	L TFG	184	−28	−40	−14	4.74
	L POC	159	−56	−30	22	4.92
	R MTG	113	58	−56	6	4.43
Delayed amount vs. K	R MTG	160	64	−48	8	4.53
	L LOC	995	−28	−72	46	5.48
	R Angular	882	40	−50	46	4.79
	L DMPFC	630	−6	16	48	5.57
	L SFG	380	−24	4	52	4.95
	R DLPFC	192	48	34	34	4.38
	R ITG	182	62	−48	−16	4.74
	R LOC	139	16	−68	60	4.23
	R MFG	125	34	4	62	4.06
	L MFG	110	−52	10	44	4.66
	R Precentral	94	50	6	24	4.31

Note: Red and blue colors represent positive and negative correlations between brain activations and K, respectively.

Abbreviations: DLPFC, dorsal lateral prefrontal cortex; DMPFC, dorsomedial prefrontal cortex; ITG, inferior temporal gyrus; LOC, lateral occipital cortex; MFG, middle frontal gyrus; MTG, middle temporal gyrus; OP, occipital pole; POC, parietal operculum cortex; SFG, superior frontal gyrus; TFG, temporal fusiform gyrus.

Delayed‐loss relevant brain activity was found to be negatively correlated with K in the prefrontal‐parietal network (Figure 3b; Table 2), including the left DMPFC (xyz = −6, 16, 48, Z = 5.57, cluster size = 630), right DLPFC (xyz = 48, 34, 34, Z = 4.38, cluster size = 192), left SFG (xyz = −24, 4, 52, Z = 4.95, cluster size = 380), right angular gyrus (xyz = 40, −50, 46, Z = 4.79, cluster size = 882), right MFG (xyz = 34, 4, 62, Z = 4.06, cluster size = 125), left MFG (xyz = −52, 10, 44, Z = 4.66, cluster size = 110). Similar negative correlation patterns were found in the left LOC (xyz = −28, −72, 46, Z = 5.48, cluster size = 995), right inferior temporal gyrus (ITG; xyz = 62, −48, −16, Z = 4.74, cluster size = 182), right LOC (xyz = 16, −68, 60, Z = 4.23, cluster size = 139), and right precentral gyrus (xyz = 50, 6, 24, Z = 4.31, cluster size = 94). However, delayed‐loss relevant brain activity in the right MTG (xyz = 64, −48, 8, Z = 4.53, cluster size = 160; Figure 3b) was positively correlated with K. Further conjunction analysis revealed that there were only two conjunction brain regions (i.e., DMPFC and DLPFC, Figure 3c) where activations were positively correlated with K in the immediate‐loss condition and negatively associated with K in the delayed‐loss condition. Given the close relationship between K and drift rate, immediate/delayed‐loss relevant brain activity was found to correlate with drift rate (Figure S1), suggesting a potential common neural substrate underlying impulsivity, even when using two distinct computational strategies (e.g., K vs. drift rate).

Drawing on these results, we hypothesized that individual impulsivity could be influenced by the human brain's computation of the subjective value of immediate and delayed‐loss options. This computation is equivalent to the processing of different options, taking into account a fixed delay time length. Specifically, if the human brain successfully processes the subjective value of immediate‐loss options, resulting in larger brain activations, particularly in the DMPFC and DLPFC under immediate‐loss conditions, individuals are more likely to select immediate‐loss options. Conversely, if the human brain effectively processes the subjective value of delayed‐loss options, individuals are more likely to choose the delayed‐loss alternatives, as evidenced by larger delayed‐loss‐related activations in the DMPFC and DLPFC. Indeed, such hypotheses have been substantiated by the aforementioned associations between K and brain activations in immediate and delayed‐loss conditions. Furthermore, we proposed that the differences in brain activations processing the subjective value of immediate and delayed loss options also play a crucial role in predicting individual K. Once again, we found a positive correlation between K and the differences in brain activations between immediate and delayed values on DMPFC_IM‐DL (r = .613, p < .001; Figure 3c) and DLPFC_IM‐DL (r = .547, p < .001; Figure 3c). This finding supports the hypothesis that the ability of individuals to successfully evaluate the value of immediate and delayed‐loss options, as well as the differences in their activation, could be a crucial neural mechanism that explains individual variations in impulsivity within the loss framework.

Of the close associations between v and K, we further tested whether the associations between brain activations, when processing the immediate and delayed loss amounts and K are mediated by the drift rate estimated by the DDM model. For the immediate‐loss relevant brain activations, v mediated the associations between K and the activations in the DMPFC (indirect effect = 0.500, the ratio between indirect effect and total effect: η = 77.28%, 95% CI [0.233 0.777]; Figure 3a) and DLPFC (indirect effect = 0.432, η = 77.14%, 95% CI [0.132 0.733]; Figure 3a). v also mediated the associations between K and delayed‐loss‐related brain activations in the DMPFC (indirect effect = −0.598, η = −84.34%, 95% CI [−0.848–0.392]; Figure 3b) and DLPFC (indirect effect = −0.666, η = −92.89%, 95% CI [−0.938–0.398]; Figure 3b). Focusing on the activation differences between the processing of the immediate and delayed loss amounts (e.g., DMPFC_IM‐DL and DLPFC_IM‐DL), v mediated the association between the brain activation differences and K (DMPFC_IM‐DL: indirect effect = 0.361, η = 78.99%, 95% CI [0.251 0.499]; DLPFC_IM‐DL: indirect effect = 0.385, η = 83.15%, 95% CI [0.221 0.540]; Figure 3c). However, similar observations were not replicated on the remaining parameters such as a, z, and T.

3.3. Individuals with similar impulsivity exhibited similar brain activation patterns

IS‐RSA offers an advantage over traditional univariate analyses in its ability to test whether individuals who exhibit similar behavioral or personality patterns also demonstrate similar neural signals, such as brain activations and morphological characteristics. We utilized this method to investigate the association between behavioral impulsivity and brain activations in response to immediate and delayed loss amounts across individuals. For immediate‐loss relevant brain activations, we found that inter‐subject variations in impulsivity K were associated with brain activation patterns in the bilateral fusiform (r = .057, Bonferroni corrected p = .019, 1000‐time permutation r = .024, corresponding p = .05)，MTG (r = .055, Bonferroni corrected p = .039, 1000‐time permutation r = .023, corresponding p = .05), the left ITG (r = .054, Bonferroni corrected p = .044, 1000‐time permutation r = .032, corresponding p = .05), TP (r = .059, Bonferroni corrected p = .011, 1000‐time permutation r = .027, corresponding p = .05), the inferior MOG (r = .075, Bonferroni corrected p = 5.972 × 10⁻⁵, 1000‐time permutation r = .031, corresponding p = .05), SMG (r = .065, Bonferroni corrected p = 1.711 × 10⁻⁴, 1000‐time permutation r = .028, corresponding p = .05) and the superior MOG (r = .062, Bonferroni corrected p = 8.331 × 10⁻³, 1000‐time permutation r = .024, corresponding p = .05) (Figure 4).

FIGURE 4.

Inter‐subject representational similarity analysis. Inter‐subject RSA revealed brain parcels where individuals with similar impulsivity exhibited similar BOLD patterns in the loss‐amount‐related processing. Behavioral dissimilarity on impulsivity K was computed by the absolute differences between each pairwise of subjects. Neural pattern dissimilarity likely reflects the potential cognitive computations underlying impulsivity under the loss framework. Brain slice numbers represent coordinates in MNI space. FOC, orbital part of middle frontal gyrus; FFG, fusiform gyrus; IOG, inferior occipital gyrus; MTG, middle temporal gyrus; MOG, middle occipital gyrus; SMG, supramarginal gyrus; IFG, inferior frontal gyrus; PreCG, precentral gyrus; IPL, inferior parietal gyrus; MFG, middle frontal gyrus; SMA, supplementary motor area; DLPFC, dorsolateral prefrontal cortex; DMPFC, dorsal medial prefrontal cortex.

We observed similar patterns for delayed‐loss relevant activation patterns in the IFG (r = .058, Bonferroni corrected p = .014, 1000‐time permutation r = .025, corresponding p = .05), left PreCG (r = .085, Bonferroni corrected p = 1.130 × 10⁻⁶, 1000‐time permutation r = .045, corresponding p = .05), IPL (r = .059, Bonferroni corrected p = .010, 1000‐time permutation r = .026, corresponding p = .05), MFG (r = .060, Bonferroni corrected p = .008, 1000‐time permutation r = .026, corresponding p = .05), fusiform (r = .061, Bonferroni corrected p = .006, 1000‐time permutation r = .057, corresponding p = .05), the left DLPFC (r = .076, Bonferroni corrected p = 3.728 × 10⁻⁵, 1000‐time permutation r = .047, corresponding p = .05) and SMA (r = .056, Bonferroni corrected p = .025, 1000‐time permutation r = .022, corresponding p = .05) (Figure 4). Moreover, we found similar patterns showing a close association with brain activation patterns (e.g., including immediate and delayed loss amount) on the drift rate from the DDM computational model on these brain regions.

Further mediation analyses also showed individual differences in the drift rate mediating the associations between inter‐subject variations in impulsivity K and in immediate‐loss‐related brain activation patterns (Figure 5a), including the bilateral fusiform (indirect effect = 0.122, the ratio between indirect effect and total effect: η = 88.41%, 95% CI = [0.060 0.187]), MTG (indirect effect = 0.119, η = 81.51%, 95% CI = [0.046 0.190]), left ITG (indirect effect = 0.063, η = 91.30%, 95% CI = [0.031 0.095]), TP (indirect effect = 0.138, η = 81.66%, 95% CI = [0.063 0.213]), inferior MOG (indirect effect = 0.133, η = 84.18%, 95% CI = [0.070 0.195]), SMG (indirect effect = 0.119, η = 98.34%, 95% CI = [0.069 0.167]) and superior MOG (indirect effect = 0.099, η = 77.95%, 95% CI = [0.046 0.154]). Likewise, similar mediation effects were observed in delayed‐loss related brain activation patterns, including the IFG (indirect effect = 0.110, η = 93.64%, 95% CI = [0.062 0.159]), left PreCG (indirect effect = 0.139, η = 97.89%, 95% CI = [0.095 0.185]), IPL (indirect effect = 0.085, η = 96.59%, 95% CI = [0.041 0.127]), MFG (indirect effect = 0.084, η = 91.30%, 95% CI = [0.043 0.127]) fusiform (indirect effect = 0.128, η = 115.92%, 95% CI = [0.078 0.179]), left DMPFC (indirect effect = 0.122, η = 91.73%, 95% CI = [0.071 0.171]), and SMA (indirect effect = 0.063, η = 84.00%, 95% CI = [0.024 0.102]) (Figure 5b).

FIGURE 5.

The relations among the inter‐subject similarities on the brain activations, fitting parameters from a computational model, and behavioral index. Individual differences in drift rate mediated the associations between immediate‐loss‐related brain activation patterns in the fusiform, MTG, TP, ITG, inf_MOG, SMG, and sup_MOG, and individual differences in impulsivity K (a). Such characteristics were further observed in delayed‐loss‐related brain activation patterns in the IFG, PreCG, MFG, DMPFC, IPL, fusiform, and SMA (b).

3.4. Neuroanatomical substrates underlying the individual differences in impulsivity

We further investigated whether GMV was associated with individual variability in impulsivity (Figure 6). We found that K was positively correlated with GMVs in the right OFC (xyz = 14, 38, −26, Z = 3.28, cluster size = 962) but negatively correlated with GMVs in the left precuneus (xyz = −16, −54, 40, Z = 3.82, cluster size = 798) and right cerebellum (xyz = 24, −72, −58, Z = 3.88, cluster size = 1588) after controlling for age, gender, parental education levels, and TIV.

FIGURE 6.

Morphological associations with impulsivity and their mediation effects. The gray matter volumes (GMVs) in the OFC (red color: positive correlation), precuneus (PCUN; blue color: Negative correlation), and cerebellum (Cereb; blue color: negative correlation) were significantly correlated with impulsivity index K. Drift rate mediated the associations between GMVs in these regions and K.

Considering the close link between v and K, we further explored whether v mediated the associations between brain morphological characteristics and K. For the OFC, v mediated GMV and K (indirect effect = 1.350, η = 91.40%, 95% CI [0.455 2.348]; Figure 6) and similar patterns were found for the GMVs in the precuneus (indirect effect = −0.564, η = 96.25%, 95% CI [−0.942–0.207]; Figure 6) and cerebellum (indirect effect = −0.569, η = 109.00%, 95% CI [−0.944–0.364]; Figure 6).

4. DISCUSSION

Our study provides a novel perspective on impulsivity by exploring its computational and neural underpinnings within a loss framework. The observed positive correlations between the drift rate and both K and the proportion of immediate‐loss options chosen align with the DDM proposition of decision‐making as a dynamic process (Ratcliff et al., 2016), not only extending prior research primarily focused on the gain domain (Amasino et al., 2019), but also suggesting that the dynamics of decision‐making could be significant factors in understanding why some individuals are more impulsive than others.

Our findings identified GMV as a notable correlate of impulsivity; the positive correlation between impulsivity and GMV in the OFC supports its known role in intertemporal choice through valuation and imagination (Kable & Glimcher, 2009; Mainen & Kepecs, 2009; McClure et al., 2004). The negative correlation with GMVs in the precuneus and cerebellum—regions implicated in various high‐level cognitive processes and delay discounting rates (Elton et al., 2017)—similarly underscores their suggested role in impulsivity.

4.1. Neural substrates underlying loss amount

Further, we uncovered several neural substrates underlying the processing of immediate and delayed‐loss amounts. For the immediate‐loss amount, brain activations in several regions, including the prefrontal‐parietal cortex (e.g., MFG, IFG, FP, and SPL) and precuneus, were modulated. In past studies investigating the gain framework, the prefrontal cortices were found to represent the amount of immediate rewards (Wang et al., 2014) and process the amount of future rewards (Ballard & Knutson, 2009; Wang, Wang, et al., 2021). The prefrontal cortex and SPL exhibit greater activity when comparing losses against gains‐related choices (Xu et al., 2009), and the prefrontal‐parietal cortex—a critical hub in the top‐down cognitive control network (Dosenbach et al., 2010)—show less activity when processing the amount information of immediate‐loss options, reflecting reduced cognitive modulations. This aligns with the preference of individuals to undergo an aversive experience immediately, a behavior that does not require cognitive control (Berns et al., 2006). Our study unveils a novel facet of neural mechanisms during loss processing, particularly the association between the cognitive appraisal of immediate loss amounts and the diminished activity in the prefrontal‐parietal cortex. This discovery profoundly enriches our comprehension of neural dynamics in decision‐making, imparting fresh neurobiological perspectives on how individuals assess immediate versus delayed options in loss scenarios. These insights may contribute significantly to our understanding of human decision‐making behavior in risky and uncertain environments, paving the way for innovative interventions that aim to refine and optimize the decision‐making process.

We found that the precuneus not only mediated the associations between its GMV and the immediate‐loss preference index (Wang, Zhang, et al., 2023) but also yielded close associations with impulsivity. This is consistent with previous research highlighting precuneus's role in processing emotional and death‐related information (Han & Kim, 2010; Kazén et al., 2012). Our findings suggest that certain death‐related subjective experiences may influence individual immediate‐loss selection, providing a new perspective on the cognitive processes underlying impulsivity (Yanagisawa et al., 2023).

In addition, brain activations in several regions were negatively modulated by the amounts of immediate and delayed‐loss options, especially those involved in visual processing and cognitive control systems. When individuals face a loss (either immediate or delayed), diminished activations in the cognitive control system may indicate less cognitive effort to process the amounts or value of immediate and delayed‐loss options, as well as their relative values. Prior studies have supported this speculation, showing that relative values between immediate and delayed rewards were negatively correlated with brain activations in the cognitive control system (e.g., including SFG and SPL) (Wang et al., 2014). Furthermore, when the magnitudes of immediate‐ and delayed‐loss options are large, the neural processing related to amount or subjective value might require fewer cognitive resources to perform the underlying psychological computations automatically. Taken together, neural activations that positively or negatively correlate with the amounts of options might be critical for understanding the potential mechanisms in loss framework.

4.2. Value‐based theory and valuation processing

The value‐based theory suggests that individuals engage in a sequence of computing the value of different options and comparing them before choosing the option with greater value (Rangel et al., 2008). This theory has widely been supported in numerous intertemporal choice studies within the gain framework (Deng et al., 2023; Kable & Glimcher, 2009; Smith et al., 2009; Wang et al., 2014). Within the loss framework, we propose that valuation processing, as dependent on the DMPFC and DLPFC, plays a central role in understanding individual differences in choosing immediate‐loss options. In other words, the computation of the loss value of immediate options, subserved by the DMPFC and DLPFC, could be a key factor influencing choice behavior. The observed correlations between impulsivity and brain activations in these regions further underscore the importance of valuation in understanding individual variability in impulsivity under the loss framework. Our findings on the DMPFC and DLPFC (Lv et al., 2019; Lv et al., 2021; Wang et al., 2016; Wang et al., 2020; Wang, Wang, et al., 2021) contribute to a growing body of research highlighting their critical roles in representing reward amounts and predicting individual decision impulsivity (Wang et al., 2016; Wang, Wang, et al., 2021). More importantly, these findings offer a unified framework to understand the cognitive and neural mechanisms underlying impulsivity across different domains. By extending these insights to the realm of loss valuation, we contribute novel perspectives to the field and provide a foundation for future investigations.

4.3. Latent cognitive process, decision‐making dynamics, and neural activation patterns

We further found that a latent cognitive process, as assessed by the DDM model, mediated the associations between valuation‐related brain activity in the DMPFC/DLPFC and behavioral impulsivity. Specifically, the dynamics of decision‐making, particularly the drift rate rather than the starting point, could be a significant factor in understanding why some individuals are more impulsive than others. Although we observed that individuals exhibited significant response biases toward the immediate‐loss options in the loss framework (starting point z; Exp. 1: t ₍₉₆₎ = 12.69, p < .001; Exp. 2: t ₍₅₇₎ = 8.74, p < .001), impulsivity was not correlated with the starting point parameters. One possible explanation is that the loss framework might induce systematic response biases toward immediate‐loss options, while impulsivity might selectively influence the process of evidence accumulation. Moreover, we observed that impulsive individuals exhibited typical perception processing before making a formal decision, consistent with the intuitive notion that impatient individuals are more prone to respond quickly. Furthermore, impulsive individuals might be influenced less from the noise. This is the first work to investigate the relationships between model parameters (e.g., starting point, drift rate, decision threshold, and nondecision time) and impulsivity index, highlighting the critical process of evidence accumulation on loss‐related intertemporal choices.

Utilizing the IS‐RSA, our results showed that individuals with similar levels of impulsivity tended to exhibit similar neural activation patterns in the visual pathway when processing immediate‐loss amounts. These patterns, possibly reflecting typical information processing, may drive latent cognitive processing and guide impulsive decisions. This suggests that individuals with similar impulsivity may exhibit similar attention selection processes when processing immediate loss amounts. In contrast, when processing delayed‐loss amounts, individuals with similar impulsivity displayed similar neural activation patterns, particularly focusing on the prefrontal cortex and motor cortex. These findings suggest that the prefrontal cortex is highly variable, as indicated by its influence on impulsivity through distinct mechanisms, such as activation strength and pattern. Our study highlights the importance of considering both immediate and delayed‐loss options in the study of impulsivity and contributes to a growing body of research highlighting the critical roles of the DMPFC and DLPFC in representing reward amounts and predicting individual decision impulsivity.

Based on the fact that most of the previous authors used rank correlation calculations and we used Pearson's calculations in our analyses, we likewise used Spearman's rank correlation calculations to validate our results, which are detailed in the Supplementary Material (Table 1), and likewise, we obtained the same results as before.

4.4. Brain structure in impulsivity

The delayed‐loss amount negatively correlated with brain activity in the visual pathway, a critical route for processing delayed‐loss amount information. Past studies have consistently documented its role in reward processing during intertemporal choice under the gain framework (Lempert & Phelps, 2016; McClure et al., 2004; Peters & Büchel, 2011). Our findings suggest this role may extend to loss processing by providing early perceptual processing for critical information. Specifically, we speculate that disrupted valuation processes for future‐loss options might drive individuals' propensity to choose immediate‐loss options without waiting. By uncovering this correlation, we contribute a novel perspective to the understanding of loss valuation. Specifically, we speculate that when the visual pathway is disrupted in its ability to effectively process information about future‐loss options, it may drive individuals toward choosing immediate‐loss options without hesitation. This finding underscores the importance of early perceptual processing in shaping individuals' loss aversion and decision‐making tendencies.

Finally, our study contributes to the understanding of impulsivity by highlighting the role of brain structure, particularly GMV in the OFC and the precuneus. Impulsivity was positively associated with GMV in the OFC, a region known for its role in intertemporal choice through valuation and imagination and implicated in the reward circuit (Kable & Glimcher, 2009; Mainen & Kepecs, 2009; McClure et al., 2004). This association suggests that the structural characteristics of the OFC may influence an individual's propensity for impulsivity in the context of immediate rewards. Past studies have likewise implicated this based on findings where lesions in the OFC have been associated with an increased preference for immediate rewards (Peters & D'Esposito, 2016; Sellitto et al., 2010). Conversely, impulsivity was negatively correlated with GMVs in the precuneus and cerebellum. Precuneus is considered a hub region within the default mode network, serving multiple functions such as introspective self‐reference (Mitchell et al., 2011), processing delay time (Weber & Huettel, 2008), encoding value differences (Vanyukov et al., 2017) within gain framework of intertemporal choice. Additionally, uncertain and irrational decisions have been shown to activate the precuneus region (Kirk et al., 2011; Krug et al., 2014; Paulus et al., 2001). Hence, under the loss framework, individuals may make impulsive decisions due to experiencing higher risk and uncertainty, altered valuation, and disruptions in time processing. Moreover, the cerebellum, known to engage in various high‐level cognitive processes such as hierarchical cognitive control (D'Mello et al., 2020), emotion regulation (Schutter & van Honk, 2009), and performance monitoring (Peterburs & Desmond, 2016), has been associated with the delay discounting rate (Elton et al., 2017), suggesting that these regions may modulate impulsivity in the context of delayed rewards. Previous studies have also demonstrated the engagement of the cerebellum in the representation of delayed/probability rewards (Wang et al., 2014; Wang, Zhang, et al., 2023) and encoding relative value (Wang et al., 2014), particularly within gain framework. Once again, our findings underscore the importance of the cerebellum in understanding impulsivity, especially within the context of the loss framework.

4.5. Limitations and future directions

Several considerations are noteworthy in this study. First, this study assessed only the state level of choice impulsivity, making it challenging to generalize the findings to trait impulsivity measured by different impulsivity‐related questionnaires. Future studies should explore the neural substrates underlying individual differences in trait impulsivity and compare them with state impulsivity. Second, the sample of this study primarily consisted of university students, limiting the generalizability to different populations. Moreover, the diversity of this sample, especially in terms of cultural background and identity, may influence the conclusions due to the close associations between the factors and impulsivity. In the future, we will further validate their potential impacts and underlying mechanisms. Third, we found significant correlations between the GMVs in several regions and impulsivity. However, it remains unclear to what extent cortical thickness and brain areas, which commonly determine the size of GMV, contributed to brain‐behavior associations. Future research should provide more insights into this topic. Finally, while inter‐subject variations in drift rate mediated the associations between inter‐subject variations in K and brain activation patterns in several regions when processing immediate‐ and delayed‐loss options, such analyses are determined by a data‐driven rather than a hypothesis‐driven strategy. One critical reason is the lack of a clear relationship between brain activation patterns and impulsivity, especially in the loss framework. Therefore, we need to be careful in our inferences based on the findings reported in the main text.

5. CONCLUSION

In conclusion, our study underscores the multifaceted nature of impulsivity, highlighting the interplay between computational and neural factors. The use of a revised paradigm of intertemporal choice within a loss framework and the novel analytic approach of IS‐RSA has revealed potential shared neural patterns among individuals with similar impulsivity levels. These findings perhaps open exciting possibilities for future research and hold promise for the development of more effective strategies to understand and manage impulsivity. Our study emphasizes the potential need for a comprehensive approach to studying impulsivity, providing a valuable foundation for future explorations in this field.

CONFLICT OF INTEREST STATEMENT

The authors declare no conflicts of interest.

Supporting information

DATA S1: Supporting Information.

Jiang, K. , Zhao, G. , Feng, Q. , Guan, S. , Im, H. , Zhang, B. , Wang, P. , Jia, X. , Zhu, H. , Zhu, Y. , Wang, H. , & Wang, Q. (2024). The computational and neural substrates of individual differences in impulsivity under loss framework. Human Brain Mapping, 45(11), e26808. 10.1002/hbm.26808

DATA AVAILABILITY STATEMENT

The data that support the findings of this study are available from the Functional MRI Center at Tianjin Normal University (TJNU). Data and code are available from the corresponding authors with the permission of the TJNU.