Neural responses to intention and benefit appraisal are critical in distinguishing gratitude and joy.

Gratitude and joy are critical for promoting well-being. However, the differences between the two emotions and corresponding neural correlates are not understood. Here we addressed these issues by eliciting the two emotions using the same stimuli in an fMRI task. In this help reception task, participants imagined them in a situation where they need financial aid. Critically, we manipulated the benefactor’s intention to provide help and the value of the benefit. Behaviorally, gratitude was stronger than joy when the benefactor-intention was strong and the benefit-value was low compared to other conditions. In parallel, gratitude activated mentalizing-related (e.g. precuneus) and reward-related regions (e.g. putamen) more strongly than joy in corresponding conditions compared to others. Moreover, gratitude was more negatively (or less positively) encoded in the region associated with mentalizing (i.e. the left superior temporal gyrus) than joy. Multivariate pattern analysis further demonstrated that the modulation patterns of benefactor-intention and benefit-value in mentalizing-related (e.g. precuneus, temporo-parietal junction) and reward-related regions (e.g. putamen, perigenual anterior cingulate/ventromedial prefrontal cortex) could distinguish the two emotions. The findings suggest that benefactor-intention and benefit-value appraisal and their neural correlates are critical in distinguishing gratitude and joy. Direct implications for gratitude interventions were discussed.


Methods
Participants. Thirty Chinese college students (15 females; 19-30 years, mean ± SD = 23.40 ± 2.98) participated in the experiment. All participants were right-handed and had normal or corrected-to-normal vision. All participants gave informed consent. The study protocol was approved by the Institutional Review Board of Tsinghua University, and the experiment was carried out in accordance with the approved guidelines.

Design and Procedure
In this study, we used scenario paradigm to induce gratitude and joy so that we were able to manipulate benefactor-intention with various levels (rather than simply intentional vs. unintentional) and compare the neural representations of benefactor-intention between gratitude and joy. Previous studies comparing emotions with imaginary scenario paradigm employed different stimuli to induce corresponding emotions 8 , so they could not rule out the possibility that the differences in neural correlates may be just out of stimuli differences. To address this issue, we designed a task with the same stimuli to induce both gratitude and joy.
The experiment has a 3 (benefactor-intention: strong, weak, no) ×3 (benefit-value: high, low, zero) ×2 (emotion: gratitude, joy) within-subject design. In each trial, the factors of benefactor-intention and benefit-value were manipulated, while the emotion factor was not manipulated but used only as a cue to which emotion (gratitude or joy) participants were asked to imagine. Specifically, during the experiment, participants were instructed to imagine how they would feel in the situation described on each screen when they were in urgent need of money. As shown in Fig. 1, in each trial, a screen with two lines of text ("The degree your friend wants to help you is X%", "who gives you ¥Y") was shown for 3 seconds, followed by a 500 ms fixation. The X% varied from 95% to 100% in the strong benefactor-intention condition (Intention Strong ), 7% to 12% in the weak benefactor-intention condition (Intention Weak ), and 0% in the no benefactor-intention condition (Intention No ). The ¥Y varied from ¥800 to ¥1200 in the high benefit-value condition (Value High ), from ¥80 to ¥120 in the low benefit-value condition (Value Low ), and ¥0 in the zero benefit-value condition (Value Zero ). On the following screen, participants rated how grateful they felt within 4-8 seconds, by pressing the button corresponding to "strongly", "moderately" or "weakly". This was followed by a 500 ms fixation. Finally, another screen instructing participants to rate how joyful they feel was shown for 4-8 seconds, also followed by a 500 ms fixation. The order of the two ratings was counterbalanced across runs in one of two orders (A-B-B-A-A-B or B-A-A-B-B-A) evenly distributed across subjects. The duration of the two ratings in each trial was one of the following permutations: 8 s/4 s, 7 s/5 s, 6 s/6 s, 5 s/7 s, 4 s/8 s. As a result, the total duration of each trial was consistently 16.5 s.
The nine conditions (3 benefactor-intention by 3 benefit-value levels) were presented in pseudo-random order to ensure the same level of benefactor-intention or benefit-value was not presented on two consecutive trials. Six runs of twenty trials, including two null trials consisting of asterisks with lengths equaling to those of the stimuli in other trials, were carried out in the task. Note that one out of the thirty participants completed only four runs of the experiment and data from trials where either gratitude or joy rating was missing were excluded from all analyses.
Neuroimaging data acquisition. Functional MRI was performed on a Philips Achieva 3.0 T TX scanner with a SENSE 8-channel head coil at the Center of Biomedical Engineering, Tsinghua University. T1-weighted structural images were acquired with TR = 8.2 ms, TE = 3.8 ms and flip angle = 8°. The SENSE factor was 2/1.5 for AP/RL, and the acquisition matrix was 256 mm × 256 mm. One hundred and sixty contiguous sagittal slices were acquired with voxel size of 1×1×1mm 3 . T2-weighted functional images were acquired with TR = 2300 ms, TE = 35 ms, and flip angle = 90°. FOV was 240 mm. Voxel size was 2.5 × 2.5 × 3.5 mm 3 . One hundred and forty-five volumes per run were acquired, with 37 slices per volume and slice gap = 1 mm.
Neuroimaging preprocessing. Imaging data were analyzed using the SPM12 software package (https:// www.fil.ion.ucl.ac.uk/spm/software/spm12/). T1 images of each participant were first manually adjusted to match the template for better registration. Motion and time correction were conducted to T2 images with Realignment and Slice-timing, respectively. T2 images were matched to T1 images with Coregistration and segmented into gray matter, white matter and cerebrospinal fluid with New Segmentation. Using DARTEL, gray matter template was built and registered to MNI template. The transformation parameters estimated during unified segmentation were applied to the T2 images with Normalized to MNI (and resampling to 3.0 mm cubic isotropic voxels). T2 images were smoothed with 6.0 mm FWHM Gaussian kernel.

Univariate activation analysis of neuroimaging data
Neural correlates of gratitude vs. joy and the interaction effects. To investigate the neural correlates of gratitude vs. joy and the interaction effects, a general linear model (GLM; Model 1) was created for each subject with eighteen predictors for the onsets of all conditions and nuisance regressors for the emotion rating order indicator and the six movement parameters. Out-of-brain voxels were removed by masking with the brain mask of SPM in first-level analyses 20 . Contrast images for each condition were computed for each subject.
To assess the neural correlates of gratitude vs. joy, contrast images of gratitude > joy and joy > gratitude were created for each subject. One-sample t-tests were conducted at the group level. A gray matter mask was created by binarizing SPM's prior probability gray matter map at the threshold of 0.2 and was applied to all the group-level analyses. All analyses were conducted across the whole brain, with an initial threshold of p uncorr < 0.001, and we reported as significant those regions which further passed a cluster-correction for multiple comparisons with p FWE < 0.05, using Monte Carlo and 3dClustSim in AFNI 18 (https://afni.nimh.nih.gov/). Clusters passing p FWE < 0.05 with an initial threshold of p uncorr < 0.005 were reported as trends.
Because we observed the emotion by benefactor-intention, the emotion by benefit-value, and the three-way interaction effects from the behavioral analyses, we defined and created contrast images corresponding to the behavioral results of the interaction effects in each subject as follows: The emotion by benefactor-intention interaction effect: Gratitude_(Intention Strong − 1/2Intention Weak − 1/2Intention No ) > Joy_(Intention Strong − 1/2Intention Weak − 1/2Intention No ) The emotion by benefit-value interaction effect: Gratitude_(1/2Value Low + 1/2Value Zero − Value High ) Figure 1. Timeline of experimental procedure. The order of the gratitude rating and the joy rating screen was counterbalanced across runs. where Value Low/Zero = 1/2(Value Low + Value Zero ), and Intention Weak/No = 1/2(Intention Weak + Intention No ). The same procedures of one-sample t-tests and voxel-/cluster-wise thresholds mentioned above were applied.
Comparison of neural representations of emotion rating, benefactor-intention and benefit-value between gratitude and joy. To compare the neural representations of emotion rating, benefactor-intention and benefit-value between gratitude and joy, we constructed another GLM (Model 2) for each subject, which included predictors for the onsets of gratitude and joy rating events, each followed by four parametrically varying modulators in the following order with serial orthogonalization: emotion rating, benefactor-intention in linear function and in quadratic function, and benefit-value. The same nuisance regressors for the emotion rating order indicator and the six movement parameters as in Model 1 were added in the model. Specifically, strong/weak/no benefactor-intention and high/low/zero benefit-value were coded as 1/0/−1, respectively. We added the modulator of benefactor-intention in quadratic function because the results from Model 1 showed that the Intention strong and the Intention No condition equally activated some mentalizing-related regions (e.g. TPJ/pSTS, precuneus; see Table S1) more strongly than the Intention Weak condition, which may be due to the fact that both strong and no benefactor-intention are more salient than a weak one in daily life. Out-of-brain voxels were removed in first-level analyses with the same method as applied in Model 1. Contrast images for each modulator were computed for each subject. Paired-samples t-tests were conducted at the group level to investigate whether neural representations of each modulator differed between gratitude and joy. Average beta values in the significant cluster (the left STG) were extracted for visualization.

Multivariate pattern analysis of neuroimaging data.
To examine whether the neural representations of benefactor-intention and benefit-value can dissociate gratitude from joy by corresponding beta patterns, we applied multivariate pattern analysis to classify the parametric modulator slope patterns in pairs (gratitude vs. joy). We used the unsmoothed beta maps corresponding to modulators of benefactor-intention in linear and in quadratic function and those of benefit-value (in linear function) generated for each participant from Model 2. For each pair of modulators, there are 30 × 2 beta maps in total. Therefore, for either gratitude or joy group, each subject has one beta map associated with corresponding modulator. The values of the voxels within the same gray matter mask mentioned above were extracted to generate a feature vector for each subject. Each voxel value in the beta map was a feature for classifying gratitude and joy groups.
Linear SVM was used to classify gratitude and joy groups based on the selected features. SVMs 21 are currently the most widely used supervised learning method. The parameter C was set at the default value (C = 1). The LIBSVM toolbox for Matlab was used to perform the linear SVM classification (http://www.csie.ntu.edu.tw/~cjlin/libsvm/) 22 . Leave-one-participant-out cross-validation was used to evaluate the performance of the classifier. In each fold, the two beta images of one subject were left out as the testing sample, and the remaining beta images were used as the training sample. Each feature was linearly scaled to the range of 0-1 across the training dataset, and the scaling parameters were also applied to scale the testing dataset. Within the training sample, a paired two-sample two-tailed t-test was applied to each feature, and the features with significant differences (p uncorr < 0.05) were retained 23,24 . It should be noted that this feature selection process was performed on the training set only to avoid over-fitting of the classifier. We trained a classifier using training set and the selected features, and tested it by identifying the category of the testing samples. This procedure was repeated 30 times so that the beta images for each participant in the sample were used as the testing sample once. The classification accuracy was calculated, which was defined as the quantity of beta images that were correctly classified. Also, specificity and sensitivity were computed, which are the proportions of gratitude and joy samples correctly classified, respectively. The permutation test was applied to determine whether the classification accuracy was significantly higher than expected by chance 23,25,26 . We applied the above prediction procedure 1,000 times, each time we permuted the labels across the training samples without replacement. The p value of the classification accuracy was the portion of the permutations that showed a higher value than the actual accuracy for the real sample.
Then, we localized the pattern information that contributed to the classification. In each fold, the feature selection was based on a slightly different subset of the data. Thus, the selected features differed slightly from fold to fold. The relevant features were restricted to those selected in every validation fold 23,24 . It is well established that the weight vector of a linear SVM classifier represents the contribution of the features to the classification 27,28 . The discriminative weight for each feature was defined as the average of their absolute weight across all folds. Specially, the absolute value of the weight quantifies the contribution of the corresponding feature to the classification. According to prior literature 29 , we reported clusters with at least 5 voxels that contributed to the prediction.
To investigate whether benefactor-intention or benefit-value has different effects on gratitude and joy, we conducted an emotion (gratitude vs. joy) by benefactor-intention by benefit-value repeated measures ANOVA with 13. The interaction effect between benefactor-intention and benefit-value was not significant, but the three-way interaction effect was: F(4, 116) = 3.35, p = 0.043, η p 2 = 0.10. Multiple comparisons showed the difference between gratitude and joy rating in the Intention Strong condition (gratitude: 2.56 ± 0.30; joy: 2.34 ± 0.36) was significant: t(29) = 4.38, p < 0.001; while those in the Intention Weak and the Intention No conditions were not (Fig. 2c). Likewise, the difference between gratitude and joy rating was significant in the Value Low condition (gratitude: 2.05 ± 0.41; joy: 1.92 ± 0.47): t(29) = 2.76, p = 0.010; and in the Value Zero condition (gratitude: 1.46 ± 0.31; joy: 1.32 ± 0.33): t(29) = 3.98, p < 0.001; but not in the Value High condition (Fig. 2d). For the three-way interaction, multiple comparisons showed that gratitude was significantly larger than joy rating in the Value Low _Intention Strong condition (gratitude: 2.64 ± 0.36; joy: 2.40 ± 0.54): t(29) = 3.76, p = 0.001; and in the Value Zero _Intention Strong condition (gratitude: 2.06 ± 0.65; joy: 1.69 ± 0.62): t(29) = 3.93, p < 0.001; while no significant difference was found in any other condition (Fig. 2e). Based on the results mentioned above, we combine results by averaging ratings across conditions of Value Low and Value Zero as well as those of Intention Weak and Intention No , and gratitude was significantly larger than joy rating only in the Value Low/ Zero _Intention Strong condition (gratitude: 2.35 ± 0.44; joy: 2.04 ± 0.54): t(29) = 4.19, p < 0.001 (Fig. 2f).

Neuroimaging results of univariate activation analysis
Neural correlates of gratitude vs. joy and the interaction effects. The neuroimaging analysis failed to show suprathreshold differences in average activation between gratitude and joy. No suprathreshold differences were observed for the emotion by benefactor-intention interaction contrast either. On the other hand, corresponding to the emotion by benefit-value interaction effect in the behavioral results, the contrast showed a significant activation in the left cuneus extending to precuneus, and trends in the precuneus extending to superior parietal lobule, the calcarine extending to precuneus and the left STG (Table 1 & Fig. 3a). Corresponding to the three-way interaction effect in the behavioral results, the contrast showed a significant activation in the left putamen extending to pallidum, the right inferior occipital gyrus (IOG) extending to fusiform/ITG, and a trend in the left dorsolateral prefrontal cortex (Table 1 & Fig. 3b). Results of contrasts between levels of benefactor-intention or benefit-value are shown in Table S1.

Neuroimaging results of multivariate pattern analysis
According to the leave-one-participant-out cross-validation, the classification accuracy for gratitude vs. joy by linear modulation pattern of benefactor-intention was 68.33% (permutation test p perm < 0.001; specificity: 66.67%; sensitivity: 70%), that by quadratic modulation pattern of benefactor-intention was 41.67% (p perm = 0.899; specificity: 33.33%; sensitivity: 50%) and that by linear modulation of benefit-value was 60% (p perm = 0.024; specificity: 73.33%; sensitivity: 46.67%). The results showed that the classification accuracies for gratitude vs. joy by linear modulation of benefactor-intention and benefit-value, but not quadratic modulation of benefactor-intention, were significantly higher than expected by chance. The weight map representing contributions of all features to the discrimination between gratitude and joy by linear modulation pattern of benefactor-intention included mentalizing-related (e.g. precuneus, superior temporal sulcus [STS]/pSTS) and reward-related regions (e.g. putamen, caudate nucleus) (  Fig. 4a). The weight map to the discrimination between gratitude and joy by linear modulation pattern of benefit-value also included mentalizing-related (e.g. precuneus, TPJ, pSTS) and reward-related regions (e.g. pgACC/VMPFC, putamen) (  Fig. 4b).

Discussion
In this study, we investigated whether and how the effects of benefactor-intention and benefit-value on gratitude and joy could be dissociated at both behavioral and neural levels. As hypothesized, benefactor-intention and benefit-value showed a dissociable effect on gratitude and joy. Specifically, gratitude was stronger than joy when the benefactor-intention was strong and the benefit-value was low/zero. From the fMRI results, low/zero  compared to high benefit-value activated the left cuneus extending to precuneus more strongly in gratitude than joy, and it additionally activated the left putamen and the right IOG extending to fusiform/ITG more strongly in gratitude than joy when the benefactor-intention was strong compared to weak/no. On the other hand, gratitude was more negatively (or less positively) encoded in the left STG than joy. Multivariate pattern analysis showed that the beta patterns of benefactor-intention and benefit-value in mentalizing-related and reward-related regions were dissociable between gratitude and joy. Our study provided evidence for the critical role of neural responses to benefactor-intention and benefit-value appraisal in differentiating gratitude from joy. As hypothesized, we found that benefactor-intention showed a dissociable effect on gratitude and joy (strong benefactor-intention induced more intense gratitude than joy), which was consistent with the notion of gratitude as an other-praising or self-transcendent emotion 4,30 and empirical findings that benefactor-intention appraisal is critical to gratitude 4,14,31,32 . At the neural level, we showed that the modulation patterns of benefactor-intention in mentalizing-related (e.g. precuneus, STS/pSTS) and reward-related regions (e.g. putamen, caudate nucleus) could dissociate gratitude from joy. Previous studies have found gratitude associated with activation or gray matter volume in precuneus/PCC and pSTS/TPJ 14,16 . These regions have been broadly reported to be associated with mentalizing 18,33 . On the other hand, the putamen is a subregion of the striatum, which has been found involved in gratitude in previous studies 11,15 and widely reported to be associated with reward processing, specifically reward prediction error 19,34 . Taken together, our findings suggest that not only mentalizing but also reward processing evoked by benefactor-intention is critical in differentiating gratitude from joy, consistent with Ortony, et al. 6 's postulations that benefactor-intention serves as a cue to both other's intention and a source of desirability for gratitude but not joy. According to the theories of gratitude as adaptation for reciprocal altruism and relationship binding 35,36 , the desirability accompanying generous benefactor-intention when gratitude arises may serve as a signal of the value of a high-quality relationship. In line with this, evidence showed that gratitude increased positive relationship ratings with the benefactor, even controlling for other positive emotions 37 .
Intriguingly, the univariate analysis failed to find suprathreshold activations associated with the emotion (gratitude/joy) by benefactor-intention interaction, but the multivariate analysis showed that the parametric modulator slopes of benefactor-intention predicted gratitude and joy beyond chance-level. These results indicated that the dissociation between gratitude and joy by benefactor-intention might lie in modulation slope patterns rather than average activations in regions associated with mentalizing and reward processing. Alternatively, the interaction between emotion and benefactor-intention might be undermined by the significant three-way interaction between emotion, benefactor-intention and benefit-value. Specifically, the results showed that when the benefit-value was low/zero, strong benefactor-intention compensated the lack of benefit-value for gratitude and induced stronger gratitude than joy. These findings suggest that benefactor-intention yields desirability especially when the benefit-value is low, which is consistent with previous postulations 6 . At the neural level, fMRI results showed that the three-way interaction activated the left putamen and the right IOG extending to fusiform/ITG. As mentioned above, the putamen is associated with reward processing 19,34 , and has been found involved in gratitude benefactorintention in mentalizing-related (e.g. precuneus, STS/pSTS) and reward-related regions (e.g. putamen, caudate nucleus) contributed to the discrimination between gratitude and joy; (b) Modulation patterns of benefit-value in mentalizing-related (e.g. precuneus, TPJ, pSTS) and reward-related regions (e.g. pgACC/VMPFC, putamen) contributed to the discrimination between gratitude and joy.
in previous studies 11,15 . The fusiform gyrus may be associated with social processing. Previous VBM study has shown that gratitude related positively to gray matter volume in the fusiform gyrus 16 , which has been found associated with face perception 38 and other social cognitions, such as decoding communicative intentions 39 .
We also found an interaction effect between emotion and benefit-value, suggesting that benefit-value also showed a dissociable effect on gratitude and joy. As Ortony, et al. 6 postulated, though both gratitude and joy are influenced by benefit-value, benefactor-intention compensates the lack of benefit-value for gratitude but not joy when the benefit-value is low. Consistent with that, we found that the emotion by benefit-value interaction activated the precuneus and another two trending clusters also including precuneus. Prior meta-analysis 18 found that precuneus was involved in almost all mentalizing tasks, and might be associated with retrieving situations encoded in memory to match them with current context for selection or inference of appropriate actions and goals. In addition to precuneus, the multivariate pattern analysis showed that gratitude and joy could be dissociated by the modulation patterns of benefit-value in the right TPJ and pSTS, which are another core areas of mentalizing and involved in intention/goal inferences 18,33 . On the other hand, gratitude and joy could also be differentiated by the modulation patterns of benefit-value in pgACC/VMPFC and putamen, which are associated with reward processing 19,40 and consistently found involved in gratitude processing [12][13][14][15]41 .
Finally, we found that gratitude was more negatively (or less positively) encoded in the left STG than joy. Previous research 9,12 found activation of STG was involved in both gratitude and joy. It was thought to be associated with understanding of social schema and mental states of others from complex social signals in eye gaze, mouth movements and body language 42 . Intriguingly, we found that gratitude showed a trend of negative representation in the region. Considering that the same region also showed a trend of activation in the emotion by benefit-value interaction contrast (i.e. gratitude activated the STG stronger than joy when the benefit-value was low/zero compared to high), it is possible that when the benefit-value is low (and hence lower level of gratitude due to the main effect), individuals process complex social signals more strongly to infer mental states of the benefactor in the gratitude than in the joy event. As a result, compared to joy, gratitude rating may elicit more consistent activation of the STG for inference of the benefactor's mental states irrespective of the intensity of the emotion. On the contrary, joy rating may elicit activation of the STG stronger when the benefactor-intention and thus the joy rating is high vs. low. These findings further implied the critical role of intention appraisal or mentalizing processes in differentiating gratitude from joy.
The findings also have implications for the potential dynamic mechanisms underlying the relationship between gratitude and joy as well as gratitude interventions. From our findings, benefactor-intention and benefit-value influence both gratitude and joy, which may lay the foundation for their mutual influence. On the other hand, the different effects of benefactor-intention and benefit-value on them and their representation difference in the left STG may offer a compensatory mechanism for individuals to feel good even when the benefit-value is low, which may in turn promote long-term SWB 1,2 . Accordingly, gratitude interventions may be most efficient in boosting SWB by focusing on situations where the benefit-value is low, in which the high desirability yielded by generous intention is most saliently signaled.
While the imaginary task demonstrated a spontaneous emotion production in the laboratory setting, its ecological validity should be examined in the future. Also, other emotions, such as feeling hurt, might be evoked during the task, though they may have been evenly distributed across conditions because of the counterbalance and pseudo-randomization. It would also be interesting to include a wider variety of measures to address the differences between gratitude and joy in the future, such as physiological measures (e.g. vagal tone) and facial expression. In addition, we did not investigate whether gratitude and joy showed different effects on subsequent behavior. As a self-transcendent emotion and one adapted for relationship binding 30,36 , gratitude should show stronger effect on affiliation and prosocial behavior than joy. Future studies should further investigate these topics, as well as the neural mechanisms underlying the dynamic relationship between gratitude and joy.

Data availability
The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request.