Sixty-two subjects (26 males; age ranged from 18 to 26 years old; Mean age ± SD = 20.98 ± 1.43) participated in our study and their structural MRI data were collected. Twenty-four subjects (8 males; age ranged from 19 to 24 years old; Mean age ± SD = 21.21 ± 1.22) performed an intertemporal choice task while their functional MRI data (fMRI) were collected. Five volunteers of this fMRI experiment task were subsequently excluded from the following analyses due to either large head motion (three subjects, mean framewise displacement [FD] > 0.5 mm in any one of three runs) or misunderstanding the task instructions (two subjects). Hence, this study included 62 structural MRI data and 19 functional MRI data. All subjects were free from neurological or psychiatric history. Informed written consent was obtained from subjects before formal experiments were conducted. This study was approved by the Institutional Review Board of the Faculty of Psychology at Tianjin Normal University (No. XL2020-27), China.
Assessment of greed personality trait
GPT was measured via a 7-item Dispositional Greed Scale (DGS)(Mussel et al. 2018; Seuntjens et al. 2015), that has been evidenced to yield high reliability and validity. Subjects indicated the extent to which they agreed with each item presented in the scale (e.g., “One can never have too much money”). All items were rated on a 5-point Likert-scale (1 = ‘strongly disagree’ to 5 = ‘strongly agree’). The total scores range from 7 to 35, whereby higher scores indicated higher levels of greedy personality.
Intertemporal choice task
Figure 1 depicts the stimuli and the experimental procedures of the intertemporal choice task. Subjects were instructed to choose between a fixed immediate reward (RMB 40—approximately USD 6) and a delayed reward that varied across trials. To estimate the neural responses to the amount and delay time of future rewards, the two dimensions were manipulated independently and orthogonally, with the amount ranging from RMB 40 to 115 (16 levels in increments of RMB 5), and the delay ranging from 1 to 150 days (16 levels in 9- or 10-day increments). These ranges were chosen based on previous studies (Wang et al., 2014) and an additional pilot study of an independent sample (n = 12). All possible combinations of each amount and delay level yielded a total of 256 trials, which were divided into three runs pseudo-randomly. We utilized an event-related fMRI design and optimized the timing and order of stimulus presentation using optseq2 in order to maximize the estimation efficiency (Dale 1999).
Following a similar procedure done in a previous study (Wang et al. 2014), for each trial, the amount and delay time of the future reward were shown on the screen but the fixed, immediate reward was told to the participant beforehand and not shown. The amount and the delay time appeared side by side divided by a vertical line (see Fig. 1) and whether the amount or the delay time appeared on the right side was determined randomly. Subjects were asked to respond as quickly as possible within the designated 3-second trial duration. If no response was made within this window, the task continued to the next trial. The trials with no responses were modeled as a separate regressor of no interest in the general linear model (GLM). To best capture subjects’ true preference for each decision, compared to a fixed, dichotomous decision rule (Tom et al. 2007), subjects indicated one of four possible responses to each decision (e.g., “strongly choose the immediate option”, “moderately choose the immediate option”, “moderately choose the delayed option”, “strongly choose the delayed option”) using a four-button response box. After each decision, the chosen option turned yellow as a feedback indication.
At the end of the experiment, all participants received the noncontingent compensation of RMB 50 plus a bonus based on the amount they actually earned on one randomly chosen trial (i.e., the participants did not receive the actual total amount earned). In other words, each participant received RMB 50 + either 40 (for choosing the immediate reward, with the total amount paid at the end of the experiment) or 45-115 (for choosing the delayed reward, with noncontingent compensation paid at the end of the experiment and the bonus paid as dictated by the delay time).
Brain imaging data acquisition
Whole-brain image data were collected using a Siemens 3T Prisma scanner with a 64-channel head coil at the Center for MRI Research of 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 in a decision task. Foam pads were utilized to minimize head motion. High-resolution T1-weighted structural images were acquired using MP-RAGE sequence and the following parameters were used: repetition time (TR) = 2530 ms; echo time (TE) = 2.98 ms; multi-band factor = 2; flip angle = 7°; field-of-view (FOV) = 224 ´ 256 mm2; slices = 192; voxel size = 0.5 ´ 0.5 ´ 1 mm3. The T2*-weighted functional images used the following parameters: TR = 2000 ms; TE = 30 ms; multi-band factor = 2; flip angle = 90˚; FOV = 224 ´ 224 mm2; slice thickness = 2 mm; slice gap = 0.3 mm; voxel size = 2 ´ 2 ´ 2 mm3. The slices were tilted ~30 degrees clockwise from the AC-PC plane to obtain better signals in the orbitofrontal cortex.
Structural MRI preprocessing and statistical analysis
Structural MRI data were collected using the Oxford Centre for Functional MRI of the Brain (FMRIB) Software Library voxel-based morphometry (FSL-VBM), a VBM style analysis toolbox implemented in FSL (version 6.00; part of the FSL package; http://www.fmrib.ox.ac.uk/fsl). Brains from the structural images were extracted, tissue-type segmented, and then aligned to the gray-matter template in the MNI152 standard space. The spatially normalized images were then averaged to create a study-specific template, to which the native gray matter images were registered again using linear and nonlinear algorithms. The registered partial volume images were then modulated by dividing them with the Jacobian of the warp field to correct for local expansion or contraction. The modulated segmented images, which represented the GMV, were then smoothed with an isotropic Gaussian kernel with 3 mm standard deviation. The smoothed data were used for the univariate analysis. In addition, we used unsmoothed GMV for further MVPA.
Firstly, we examined associations between GPT and GMV in whole-brain level using a mixed-effects FLAME 1 model implemented in FSL. Maternal education, paternal education, age at MRI scan, and total GMV were included as covariates. These factors were taken into account because behavioral correlation analysis revealed that maternal and paternal education might account for relatively large variations of GPT, although their correlations were not significant. In regression analysis, covariates were entered into the first block of equations. In the second block, mean-centered GPT and gender were entered. The interaction term, the product of mean-centered GPT and gender, were entered into the third block. When the interactive effect was not significant, a reduced model, controlling for the same covariates and gender, examined GPT in relation to the same outcome measures. Statistical results were determined at a cluster level (z > 2.3, p < 0.01) and at a family-wise error rate of 0.05 for the correction for multiple comparisons using Gaussian Random Field Theory.
Secondly, in the multivariate pattern analysis, Epsilon-insensitive support vector regression (SVR) (Drucker et al. 1997) with a linear kernel, as implemented in PyMVPA (http://www.pymvpa.org) (Hanke et al. 2009), was used to examine the associations between the GPT and GMV in whole-brain level. A searchlight procedure with a three-voxel radius (Kriegeskorte et al. 2006) was employed to provide a measure of decoding accuracy in the neighborhood of each voxel. Given previous studies (Wang et al. 2014; Jimura and Poldrack 2012), we set the e parameter in the SVR to be 0.01. A three-fold cross-validation was applied in this study. The 62 subjects were divided into three groups of 20 or 21 subjects, with matched gender as well as matched GPT. In each iteration, an SVR model was trained based on two groups of 41 or 42 subjects. Once trained, this SVR model then was applied to test the generalization on the remaining one group based on their imaging data. It is worth noting that the training dataset was firstly normalized (i.e., mean substracted out and then divided by SD) and then applied to testing dataset. The voxel-wise accuracy of SVR prediction was then calculated as the Pearson’s correlation coefficient between actual and predicted values of the GPT and then transformed to the corresponding z-score maps. Finally, SVR predictions were thresholded using cluster detection statistics, with a height threshold of z > 2.3, and a cluster probability of p < 0.05, corrected for whole-brain multiple comparisons using Gaussian Random Field Theory.
To further probe the direction of the associations between GMV and GPT, we selected the clusters showing significant prediction accuracy of GPT as ROIs. The averaged GMV were then extracted and performed a Pearson’s correlational analysis between ROI’s GMVs and GPT. To avoid the double dipping issue, we only reported the positive and negative direction of correlation, but not the exact r or p value. In addition, we extracted the z-score of ROIs showing significant prediction in MVPA respectively from the univariate and multivariate analysis in order to directly compare the effect of univariate and multivariate statistics via Pearson’s correlational analysis.
Functional MRI preprocessing and statistical analysis
The FMRI Expert Analysis Tool was used to perform the functional image preprocessing and statistical analyses. The scanner allowed discarding the first four volumes for T1 equilibrium before the task. The remaining images were then realigned to correct for head movements. Data were spatially smoothed by using a 5 mm full width at half maximum Gaussian kernel and filtered in the temporal domain using a nonlinear high-pass filter with a 90 s cutoff. EPI images were first registered to the MPRAGE structural images and then into MNI standard space, using affine transformations (Jenkinson and Smith 2001). Registration from MPRAGE structural images to standard space was further refined using FNIRT nonlinear registration (Anderson et al. 2007). Statistical analyses were firstly performed in the native image space, with the statistical maps normalized to the standard space before higher-level analysis.
The data were modeled at the first level using a general linear model within FSL’s FILM module. Five parametric regressors were included during the decision-making period starting from presentation of inter-temporal alternatives and ending when subjects responded: (1) the overall task regressor (1 for each trial); (2) the amount of delayed reward; (3) the time of delayed reward; (4) the relative value, calculated using the following formula: relative value = abs(immediate – delayed amount ´ k) (FitzGerald et al. 2009; Lim et al. 2011); (5) reaction time (RT) (Sripada et al. 2011). For this model, each regressor (except for the task regressor) was first demeaned and normalized to the same range (-1 vs 1) and then convolved with the double-gamma canonical hemodynamic response function. Trials with no valid response were modeled as a separate regressor of no interest.
A second-level analysis was performed using a fixed-effect model where all three functional runs were combined within individual participants. Finally, these contrast results were then fed into a random-effect model for group analysis and regression analysis for each individual’s GPT scores using a FLAME1 model. Group images were thresholded using cluster detection statistics, with a height threshold of z > 2.3 and a cluster probability of p < 0.05, corrected for whole-brain multiple comparisons using Gaussian Random Field Theory.