Delay discounting is associated with the fractional amplitude of low-frequency fluctuations and resting-state functional connectivity in late adolescence

As a component of self-regulation, delay discounting (DD) refers to an individual’s tendency to prefer smaller-but-sooner rewards over larger-but-later rewards and plays an essential role in many aspects of human behavior. Although numerous studies have examined the neural underpinnings of DD in adults, there are far fewer studies focusing on the neurobiological correlates underlying DD in adolescents. Here, we investigated the associations between individual differences in DD and the fractional amplitude of low-frequency fluctuations (fALFF) and resting-state functional connectivity (RSFC) in 228 high school students using resting-state functional magnetic resonance imaging (RS-fMRI). At the regional level, we found an association between higher DD and greater fALFF in the dorsal anterior cingulate cortex (dACC), which is involved in conflict monitoring and strategy adaptation. At the connectivity level, DD was positively correlated with the RSFC between the dACC and the left dorsolateral prefrontal cortex (DLPFC), a critical functional circuit in the cognitive control network. Furthermore, these effects persisted even after adjusting for the influences of general intelligence and trait impulsivity. Overall, this study reveals the fALFF and RSFC as the functional brain basis of DD in late adolescents, aiding to strengthen and corroborate our understanding of the neural underpinnings of DD.


Results
Identification of the neural basis underlying DD. Table 1 details the descriptive statistics for age, DD, general intelligence and trait impulsivity. According to the conventions 48 , the scores of each behavioral measure may be normally distributed, with skewness and kurtosis values ranging from -0.46 to 0.32. To test whether the scores of behavioral measures were normally distributed, we performed one-sample Kolmogorov-Smirnov test (K-S) 49 First, to identify the brain regions related to DD, we carried out whole-brain correlation analyses between voxel-wise fALFF values and DD scores, controlling for gender, age and head motion (i.e., framewise displacement, FD) 51 . We found a significant association between higher DD and greater fALFF in the dACC (r = 0.29, p = 0.00001; see Fig. 1 and Table 2), after correcting for multiple comparisons (Monte Carlo simulation). No other significant results were obtained in these analyses. To evaluate the stability of the relationship between DD and spontaneous brain activity, we performed prediction analyses using the linear regression and four-fold balanced cross-validation procedure. The fALFF in the dACC significantly predicted individual differences in DD [r (predicted, observed) = 0.25, p = 0.0001], even after adjusting for gender, age and FD.
Previous studies have showed that the nuisance regressors (e.g., head motion and non-gray matter signals) in the preprocessing of RS-fMRI data may affect the fALFF in the brain 52 . Thus, we reprocessed our data without the step of nuisance regression and retested the correlation between DD and fALFF. The results revealed that . We found no other significant results in these analyses. In summary, the significant region was almost the same as that identified in our initial analyses, although the size of the cluster was reduced. Thus, we used only the region detected in the initial analyses in the following analyses. Second, to further explore the role of dACC in DD, we carried out a seed-based RSFC analysis by using the dACC with significant association with DD as seed ROI and studying its connection to the rest of the brain. Then, we performed whole-brain correlation analyses between the RSFC and DD scores with gender, age and FD as controlling variables. The results revealed that after correcting for multiple comparisons (Monte Carlo simulation), DD was positively related to RSFC strength between the dACC and left DLPFC (the middle frontal gyrus; r = 0.26, p = 0.00008; see Fig. 2 and Table 2). No other significant results were obtained in these analyses. Then, to assess the stability of the association between DD and RSFC, we performed prediction analyses using the linear regression and four-fold balanced cross-validation procedure. The strength of the dACC-DLPFC connectivity significantly predicted individual differences in DD [r (predicted, observed) = 0.22, p = 0.0007], even after adjusting for gender, age and FD.
The DD-specific nature of the findings. To test the specificity of the associations between DD and intrinsic brain activity, we excluded the confounding factors of general intelligence and trait impulsivity. DD was negatively correlated with general intelligence [r = -0.14, p = 0.042 (uncorrected), p = 0.294 (Bonferroni corrected)] and positively correlated with trait impulsivity [r = 0.19, p = 0.005 (uncorrected), p = 0.035 (Bonferroni corrected)]. Then, we conducted correlation analyses to examine whether general intelligence and trait impulsivity can affect the relationships between DD and fALFF and RSFC. After controlling for general intelligence and trait impulsivity, DD was still related to the fALFF of the dACC [r = 0.28, p = 0.00002 (uncorrected), p = 0.0001 (Bonferroni corrected)] and the strength of the dACC-DLPFC connectivity [r = 0.26, p = 0.00009 (uncorrected), p = 0.0007 (Bonferroni corrected)], suggesting that the observed associations were specific to DD. Gender, age and FD were controlled for in these analyses.
Next, we performed hierarchical regression analyses to evaluate how much of the variance in DD can be explained by intrinsic brain activity. The results showed that the fALFF in the dACC [β = 0.29, p = 0.000003 (uncorrected), p = 0.00002 (Bonferroni corrected)] and the dACC-DLPFC connectivity [β = 0.26, p = 0.00003 (uncorrected), p = 0.0002 (Bonferroni corrected)] jointly accounted for 14.6% of the variance in DD (ΔR 2 = 0.146) beyond the variance explained by general intelligence and trait impulsivity as well as gender, age and FD. These results indicated that the fALFF in the dACC and the dACC-DLPFC connectivity uniquely predict individual differences in DD.

Discussion
In the current study, we sought to examine the functional brain correlates of DD in late adolescents by performing RS-fMRI. At the regional level, we found an association between higher DD and greater fALFF in the dACC. At the connectivity level, higher DD was related to stronger RSFC between the dACC and the left DLPFC. Furthermore, the fALFF in the dACC and dACC-DLPFC connectivity uniquely predicted individual differences in DD. These results persisted even after adjusting for the influences of general intelligence and trait impulsivity, indicating the DD-specific nature of the findings. In brief, the present study reveals that the regional fALFF and RSFC serve as the functional neural basis of DD in late adolescents, which aids to strengthen and corroborate our understanding of the neural underpinnings of individual differences in DD.
Confirming our first hypothesis, the fALFF in the dACC predicted individual differences in DD. This result fits well with those of previous fMRI studies revealing brain activity in the dACC during completing DD tasks 38,[53][54][55][56] . The activity of the dACC might reflect its function in conflict monitoring and strategy adaptation, which are considered key mechanisms for biasing future behaviors toward more efficient modes 14,57 . Moreover, evidence from two structural MRI studies has demonstrated that structural variations in the dACC play a critical role in individual differences in DD 17,18 . Furthermore, the result of the association between higher DD (worse self-regulation) and greater fALFF is consistent with a magnetic resonance spectroscopy (MRS) study that revealed a negative association between patient behaviors and dACC glutamate concentrations at the resting state 36 . Our results were also consistent with a body of studies reporting higher (f)ALFF in the dACC among patients with self-regulation-related disorders such as obsessive-compulsive disorder 58,59 , schizophrenia 60 , depressive disorder 61,62 , and post-traumatic stress disorder 63 . Higher fALFF values in the dACC among impatient participants and patients with self-regulation disorders might reflect an enhanced cortical modulation of neural activities 64,65 or a compensatory mechanism to overcome defects in brain function and structure 63,66 .
Confirming our second hypothesis, our study revealed an association between dACC-DLPFC connectivity and individual differences in DD, which was in line with prior functional and structural findings on DD in the brain. On one hand, DLPFC activities during DD tasks were repeatedly reported in previous studies, which demonstrated the role of DLPFC in exerting self-control to obtain greater long-term benefits 38,55,56,[67][68][69] . On the other hand, the gray matter structure of the DLPFC has been found to be linked with individual differences in DD, supporting the DLPFC as a neuroanatomical marker for DD 16,18,19 . In addition, using repetitive transcranial magnetic stimulation (rTMS), Figner et al. (2010) reported that transient disruption of the left DLPFC caused participants to select more immediate rewards, providing direct evidence for the causal role of the left DLPFC in DD 70 . Considering that the DLPFC and the dACC are generally considered the core brain regions in human self-regulation system 1, 2 , our finding regarding the association between DD and dACC-DLPFC connectivity supports the role of the cognitive control network in DD to a certain degree 14,15 .
Several limitations of this study deserve consideration in future research. First, we measured DD by using the Kirby questionnaire, which is a relatively antiquated and crude measure of DD since it only produces measures in certain bands of discounting. Future studies are encouraged to use other dynamic adjusting procedures to measure DD 38,71 , which may improve the reliability and validity of the measurement. In addition, the monetary reward for MCQ used in this study was hypothetical, although some studies have showed that the discount rates obtained using hypothetical choices are not substantially different from those obtained using real-monetary choices [72][73][74][75] . Future investigations could consider using real-monetary reward to measure DD and explore its relations to fALFF and RSFC. Second, the participants of the current study included a group of healthy high school students with a narrow age range, which may limit the generalizability of the findings, although it has the advantage of obtaining sufficient statistical power for the whole-brain analyses. Future studies are necessary to extend our study to include more diverse populations, such as the elderly, adults, children and individuals with psychiatric illnesses. Third, only fALFF and RSFC were used as measures of brain function to examine the neurobiological basis of DD during the resting state. Future studies could consider examining this issue by utilizing other measures of brain function (e.g., task-based brain activity) and structure (e.g., cortical gray matter volume or cortical thickness) and then compare the results across the different brain measures. Finally, the r values of the correlations between DD and general intelligence and trait impulsivity were rather low, which may represent the spurious correlations in considering of the large sample size. These low correlations may caused by the self-report measures used in the current study. For example, evidence from previous studies using self-reported measures has suggested that the relationship between DD behavior and impulsivity is not yet entirely clear 6 . Thus, future investigations are encouraged to use more reliable and valid measures (e.g., behavioral instead of self-reported) to examine the associations between DD, general intelligence and impulsivity.
In conclusion, the present study used RS-fMRI to explore the functional brain substrates of DD in late adolescents. The whole-brain correlation analyses suggested that individual differences in DD were positively related to the fALFF in the dACC and the RSFC between the dACC and left DLPFC. These results remained significant even after adjusting for the effects of general intelligence and trait impulsivity, demonstrating the stable and specific characteristics of our findings. In short, our study provides the evidence of the functional neural basis of DD in late adolescents. Finally, this study might have educational implications, as we provide potential neurobiological markers that can be used by education experts to develop corresponding intervening programs to promote effective decision-making and well-being in adolescents. Moreover, our study may add to Psychoradiology (https:// radiopaedia.org/articles/psychoradiology), which is an emerging subspecialty of radiology with growing intersection between the fields of clinical imaging and psychiatry/psychology 76,77 .

Methods
Participants. The participants included 234 right-handed and healthy adolescent students (mean age = 18.60 ± 0.78 years, 122 females), who were part of a longitudinal project that aimed to explore the determinants of social cognition, academic success and well-being among adolescents in Chengdu, China 31, 78-80 . Each of the students had recently graduated (by June 2015) from one of several local public high schools and did not have a history of psychiatric or neurological illness. The experiments were carried out between June 2015 and September 2015; and the Edinburgh Handedness Inventory 81 was used to measure handedness. Six participants were excluded due to abnormal brain structure (3) or a lack of behavioral test scores (3). Thus, 228 participants (mean age = 18.48 ± 0.55 years, 119 females) were included in the data analyses. The current study was approved by the local research ethics committee of West China Hospital of Sichuan University. Written informed consent was obtained from each participant prior to experimentation. After completing all the measurements, each participant received ¥100 for compensation. The study protocols were performed in accordance with the approved guidelines and regulations.

Behavioral measures. Monetary Choice Questionnaire (MCQ).
We used the 27-item MCQ to measure individual differences in DD 4 . According to the delayed reward magnitudes, the 27 items were grouped into 3 conditions: large (¥75-85¥), medium (¥50-60¥) and small (¥25-35¥), with 9 items per condition. The delay time ranged from 7 days to 186 days. For each item, the participants were asked to choose either a larger, delayed reward or a smaller, immediate reward. For instance, "Would you prefer ¥75 in 20 days or ¥41 today?" Previous evidence has shown that a hyperbolic function fits well with the responses of the participants: V = A/(1 + kD), where A refers to the delayed reward, V refers to the present reward, D refers to the delay time, and k refers to the discount rate parameter 82 . Based on a procedure developed in previous studies 3, 4 , we computed the scores for DD utilizing the following steps. First, we obtained the k value for a given delayed reward condition according to the highest consistency among a participant's choices. Second, we calculated the geometric mean of the k values of the three delayed reward conditions and then obtained a single k value for each participant. Higher k values represented higher impulsivity (i.e., more likely to select the immediate reward). Finally, we used a natural log transformation to normalize the k values (lnk) because the raw k values were not normally distributed. Evidence from previous studies has indicated that the MCQ shows good reliability and validity among adolescents and adults 4,21,83 . The MCQ has also been widely used in Chinese populations 84,85 . To evaluate the internal reliability of the MCQ, the consistency value for each participant (i.e., the percent of the consistent responses) was first calculated and then the average consistency value for all participants was obtained. This average consistency value has been employed in previous investigations 83 . In our dataset, the average consistency values were as follows: for the large delayed reward condition, mean = 98.64%, standard deviation (SD) = 3.65%, minimum = 88.89%, maximum = 100%; for the medium delayed reward condition, mean = 98.73%, SD = 3.69%, minimum = 77.78%, maximum = 100%; for the small delayed reward condition, mean = 98.73%, SD = 3.54%, minimum = 88.89%, maximum = 100%. These high consistency values suggested that the participants made their responses very carefully during testing. Because participants were compensated irrespective of their choices on the questionnaire, we used hypothetical monetary reward in this study.

Raven's Advanced Progressive Matrix (RAPM).
Because general intelligence is found to be associated with DD 11,86 and intrinsic brain activity 87 , we employed RAPM 88 to rule out the possible influences of general intelligence on the associations between DD and intrinsic brain activity. RAPM is one of the most popular and sound instruments for assessing general intelligence and includes 36 non-verbal graphical matrices. During testing, the participants were instructed to identify the missing parts for all items within 30 minutes 89 . The RAPM score was defined as the number of correct answers; the higher the score, the higher the level of general intelligence. In our dataset, the Cronbach's α value of RAPM was 0.82, suggesting an adequate internal consistency.
Barratt Impulsivity Scale-11 . Considering the associations between trait impulsivity and DD 4, 86 and intrinsic brain activity 90 , we used the 30-item BIS-11 91 to exclude the possible effects of trait impulsivity on the associations between DD and intrinsic brain activity. The BIS-11 is a widely used self-report questionnaire for assessing trait impulsivity. The scale includes three subscales: non-planning impulsivity, attentional impulsivity and motor impulsivity. The response option for items ranges from 1 (rarely/never) to 4 (almost always/always). The total score for BIS-11 was computed by summing the scores across all of the items, with higher scores reflecting greater impulsivity. A systematic review of the BIS-11 has reported that the scale exhibits strong reliability and validity across samples from different countries 92 . The Chinese versions of BIS-11 have demonstrated satisfactory psychometric properties among Chinese high school students 93,94 . In our dataset, Cronbach's α value of BIS-11 was 0.79, showing an adequate internal consistency.
Data preprocessing. Prior to preprocessing, a medical radiologist who was blind to this research visually inspected the image data for each participant. Three participants were excluded because of abnormal brain structure (e.g., unusual cyst). To ensure the stability of the MRI signals during adaption in participants, we discarded the first 10 images. Then, the remaining data were preprocessed using the following steps: slice timing and head motion correction, realignment, normalizing with 3 × 3 × 3 mm 3 resolution, smoothing using an 8 mm FWHM Gaussian kernel and removing linear trends. To remove the influences of nuisance covariates in the fALFF and RSFC analyses, we regressed out six head motion parameters 95,96 , the global mean signal, the white matter signal, and the cerebrospinal fluid signal. Finally, the data were filtered with a temporal band-pass filter (0.01 -0.08 Hz, for RSFC except for fALFF) 97 . None of the data were excluded during preprocessing for two reasons. First, the translational and rotational parameters for all of the participants did not exceed ± 1.5 mm and ± 1.5°, respectively. Second, the FD values of participants did not exceed 0.30 and these values were treated as a covariate in the brain-behavior correlation analyses. The preprocessing was performed using SPM8 (http://www.fil.ion.ucl.ac.uk/ spm/software/SPM8), which was employed using the DPARSF toolbox 98 . fALFF calculation. We calculated the fALFF based on the procedure developed by previous studies 27,28 . First, to obtain the power spectrum, we transformed the time courses of each voxel into the frequency domain. Then, after computing the square root of each frequency in the power spectrum, we obtained the mean square root across a low-frequency range (0.01-0.08 Hz). This mean square root is regarded as the ALFF index 99 . Considering that fALFF is a normalized score of ALFF, we calculated the fALFF as a fraction of the sum of the amplitudes across the entire frequency range (0-0.25 Hz). Finally, by subtracting the global mean fALFF and dividing by the standard deviation, we transformed the fALFF map into the fALFF Z-score map for each participant. These analyses were conducted using the DPARSF toolbox 98 .

Statistical analyses. fALFF-behavior correlation analyses.
To identify brain regions with spontaneous activity associated with DD, we conducted whole-brain correlation analyses between the MCQ scores and the fALFF values in each voxel, with age, gender and FD as nuisance covariates. We corrected for multiple comparisons using the AlphaSim program in the REST software package 100 , which employs Monte Carlo simulation 101 . Specifically, the threshold for significant clusters was set as follows: 10,000 iterations, p < 0.05 at the cluster level combined with a p < 0.005 at the voxel level, at least 50 voxels (1350 mm 3 ). AlphaSim is a popular program used in previous studies to analyze RS-MRI data [29][30][31]35 .

RSFC-behavior correlation analyses.
To investigate the relationship of DD and RSFC between the clusters identified from the fALFF-behavior correlation analyses and other clusters across the brain, we carried out RSFC-behavior correlation analyses using REST software 100 . First, we used the clusters with significant associations with DD to create the seed ROIs. Second, we extracted the mean time series from voxels in each seed ROI in each participant. Third, to obtain the participant-level correlation maps, we correlated the mean time series in each seed ROI with that of other voxels across the brain. Fourth, using Fisher's r-to-z transformation, we converted the correlation maps to Z-score maps. For the group-level analyses, we conducted correlation analyses between the Z-score maps and MCQ scores to detect the RSFC related to DD, with age, gender and FD as nuisance covariates. To correct for multiple comparisons, we applied the AlphaSim program and set the threshold for significant clusters as follows: 10,000 iterations, p < 0.05 at the cluster level combined with a p < 0.005 at the voxel level, at least 50 voxels (1350 mm 3 ).
Prediction analyses. We carried out a machine learning approach to test the stability of the association between DD and intrinsic brain activity. This approach is based on balanced cross-validation using linear regression, which has been widely used in previous studies 78,[102][103][104][105] . In this analysis, DD served as the dependent variable and either fALFF or RSFC was used as the independent variable. The predictive ability of independent variable on dependent variable was defined as r (predicted, observed) , which was evaluated using a four-fold balanced cross-validation procedure. First, we divided the data by four to ensure that the independent variable and dependent variable distributions across the four divisions were balanced. Second, we used three divisions to build a linear regression model, leaving out the fourth division. Then, we employed this model to predict the data for the fourth division. We repeated this procedure four times to obtain a final r (predicted, observed) , which represented the association between the observed data and the data predicted by the regression model. Here, we applied a nonparametric testing method to determine the statistical significance of the model. Specifically, The empirical null distribution of r (predicted, observed) was estimated by generating 1,000 surrogate datasets under the null hypothesis that there was no association between DD and intrinsic brain activity. We generated each surrogate dataset Di of size equal to the observed dataset by permuting the labels on the observed data points 78, 102-105 . Next, we used the predicted labels with the four-fold balanced cross-validation procedure and the actual D i labels to calculate the r (predicted, observed) of D i (i.e., r (predicted, observed)i ). Finally, we counted the number of r (predicted, observed)i values that were greater than r (predicted, observed) and divided that count by the number of D i datasets (1000). The resulting value was considered the level of statistical significance (p-value). Age, gender and FD were controlled for in these analyses.

Hierarchical regression analyses.
To examine whether the intrinsic brain activity can explain additional variance when predicting DD beyond other predictors (i.e., general intelligence and trait impulsivity) and demographic factors (i.e., age and gender), we carried out a hierarchical regression analysis using SPSS software (version 22.0). In this analysis, the dependent variable was DD; the independent variables in step 1 were general intelligence, trait impulsivity, age, gender and FD; and the independent variables in step 2 were the fALFF of brain region(s) and the RSFC identified in the prior whole-brain analyses.