key: cord-0312100-krwlvb3i
authors: Sellitto, Manuela; Neufang, Susanne; Schweda, Adam; Weber, Bernd; Kalenscher, Tobias
title: Arbitration between insula and temporoparietal junction subserves framing-induced boosts in generosity during social discounting
date: 2020-11-06
journal: bioRxiv
DOI: 10.1101/841338
sha: d04fd07e5539d86de31470d35bbad909cb11bb24
doc_id: 312100
cord_uid: krwlvb3i

Generosity toward others declines across the perceived social distance to them. Here, participants chose between selfish and costly generous options in two conditions: in the gain frame, a generous choice yielded a gain to the other; in the loss frame, it entailed preventing the loss of a previous endowment to the other. Social discounting was reduced in the loss compared to the gain frame, implying increased generosity toward strangers. Using neuroimaging tools, we found that while the temporoparietal junction (TPJ) and the ventromedial prefrontal cortex (VMPFC) subserved generosity in the gain frame, the insular cortex was selectively recruited during generous choices in the loss frame. We provide support for a network-model according to which TPJ and insula differentially promote generosity by modulating value signals in the VMPFC in a frame-dependent fashion. These results extend our understanding of the insula role in nudging prosocial behavior in humans.

Most human societies are collaborative. Collaboration offers benefits to their members that 43 they would not be able to achieve individually. However, societies can only function efficiently 44 if their members are willing to contribute to causes whose beneficiaries are abstract and 45 anonymous, such as public goods, and/or to causes whose beneficiaries are socially remote, 46 as it is often the case with wealth redistribution for social welfare, public health insurance, or state pension systems (see also Kalenscher, 2014) . Most people are indeed willing to sacrifice the gain frame, a generous choice would imply a gain of €75 to the other, while in the loss 155 frame, a generous choice would imply preventing the loss of the previous €75 endowment.

Importantly, participants were repeatedly instructed that the other person was unaware of her compatible; participants (n = 52) were paid a fixed allowance of €8.5. In the social discounting 181 task, participants can either make a selfish choice or a generous choice, in each frame 182 condition. To reconstruct the individual social discount functions, separately for the two frame 183 conditions, we fit a standard hyperbolic model (see Eq. 1 (Jones & Rachlin, 2006; Strombach 184 et al., 2015) ; Materials and methods) to trial-by-trial binary choices (i.e., either selfish or 185 generous) via a softmax function to estimate the parameter k, a measure of the steepness of 186 the social discount function. Additionally, we determined, for each participant and each social 187 distance level, and separately for the two frame conditions, the point at which the participant 188 was indifferent between the selfish and the generous alternative using logistic regression 189 (Strombach et al., 2015) . The difference in reward magnitudes for the participant between the 

To test our first hypothesis, we aimed to replicate our previous finding of brain structures 275 known to represent vicarious reward value and overcoming egoism bias in the gain frame (A. 276 Soutschek et al., 2016; Strombach et al., 2015) . Our results (GLM1; see Materials and 277 methods) indeed revealed clusters located in VMPFC (0, 54, 14, whole-brain pFWE-corr < 0.001) 278 as well as right TPJ (rTPJ; 50, -66, 36, whole-brain pFWE-corr < 0.035) to be selectively activated, 279 in addition to other prefrontal regions, when participants made generous choices in the gain 280 frame relative to generous choices in the loss frame. ROI analyses confirmed significant 281 clusters of activation in both VMPFC (pFWE-corr < 0.000) and rTPJ (pFWE-corr = 0.01). Thus, 282 consistent with (Hutcherson et al., 2015; Strombach et al., 2015) , a network comprising 283 VMPFC and rTPJ seems to underlie the motivation for costly generosity in the gain frame (Fig. 

Our second hypothesis predicted that generosity in the loss frame was motivated by social 296 norm compliance rather than other-regarding considerations; generosity should, 297 consequently, go along with a different neural activation pattern in the loss than the gain frame.

In a first step, we attempted to isolate frame-dependent neural correlates, independent of 299 participants' choices. To this end, we searched for differential neural activity at trial onset, i.e.,

when participants learned about the social distance level of the other person and which frame 301 was relevant in the current trial (see Fig. 1 ), by contrasting neural activity between the two 302 frames (GLM2; see Materials and methods). We found significant activation in the right 303 posterior insula (34, -16, 8, whole-brain pFWE-corr = 0.007) in the loss vs. gain frame contrast, 304 which was accompanied by significant activations in frontal regions, including VMPFC (2, 50,

-8, whole-brain pFWE-corr = 0.001), as well as temporal regions ( Fig. 4 ; see supplemental Table   306 S2 for a complete list of activations). ROI analyses confirmed significant clusters of activation gain frame vs. loss frame, did not reveal any significant activation. Social distance information, these contrasts (GLM2; see Materials and methods), suggesting that the activations in insula 311 and VMPFC reflected frame-but not social distance information. 

In support of this conclusion, we found that right anterior insula (42, 4, -4 , ROI analysis, pFWE-320 corr < 0.02; GLM1, see Materials and methods), was selectively activated during generous 321 choices in the loss frame relative to generous choices in the gain frame ( Fig. 5 ; see 322 supplemental Table S1 ). The location within the insula mask was slightly anterior to the peak 323 activation we found at trial onset. 

To isolate the specific functional contributions of these anterior and posterior clusters within 332 insula, we quantified the extent at which both insula spots correlated with the individual 333 propensity to make generous choices in the gain and the loss frame. To this end, we extracted 334 the parameter estimates at the individual level from an ROI with a seed in both insula clusters 335 (coordinates for the right posterior insula obtained from the results of GLM2, see above [34, -336 16, 8] , and for the anterior insula obtained from the results of GLM1, see above [42, 4, -4] ).

Parameter estimates were extracted separately for the gain and the loss frame, after pooling 338 all choices (i.e., generous and selfish) within each frame (GLM3; see Materials and methods).

We then correlated, for each frame separately, the extracted beta estimates with the 340 percentage of generous choices in the respective frame. Both insula activations positively and 341 significantly covaried with generous choices in the loss frame (posterior insula: r = 0.37, p = 342 0.02; anterior insula: r = 0.33, p = 0.03, one-tailed, medium effect size; Fig. 6a ), but not in the 343 gain frame (posterior insula: r = 0.07, p = 0.36; anterior insula: r = 0.09, p = 0.31, one-tailed; 

Our analysis so far suggests that insula activation reflects the psychological motives 355 underlying generous choice in the loss frame. However, other explanations of our insula 356 finding are conceivable, too. For instance, participants made more generous choices overall 357 in the loss than the gain frame; i.e., they forewent more own-payoff in the loss than the gain 358 frame, and insula activation might reflect the higher level of reward foregone in the loss frame.

Yet, the trial-by-trial regressor of reward amount foregone (GLM1; see Materials and methods) 360 revealed no parametric modulation of insula activity, nor of activity in any other brain region, during generous choices in either frame condition. Additionally, insula activity is unlikely to reflect the own-reward component of the generous alternative because it was fixed (always

We previously provided empirical support for a network model according to which, in a task 367 similar to our gain frame condition, TPJ would facilitate generous decision-making by 368 modulating basic reward signals in the VMPFC, incorporating other-regarding preferences into 369 an otherwise exclusive own-reward value representation, thus computing the vicarious value 370 of a reward to others (Strombach et al., 2015) . Here, we expand on this idea and propose that, 371 in addition to the TPJ-VMPFC connectivity in the gain frame, frame-related information in the 372 loss frame would activate insula, which in turn would down-regulate own-value 373 representations in VMPFC, thus promoting generous choices by decreasing the attractiveness 374 of own-rewards. Hence, in brief, we predicted a complex, frame-dependent pattern of 375 connectivity between insula, TPJ, and VMPFC that reflects the different motives underlying 376 generosity in the gain and the loss frame.

To identify the relations between those regions, we estimated their effective connectivity via 378 dynamic causal modeling (DCM) (Friston, Harrison, & Penny, 2003) . More specifically, we 379 tested the idea that the frame information at the beginning of each trial would drive increased 380 insula activation selectively in the loss frame, and increased TPJ activation selectively in the 381 gain frame. Additionally, we expected increased endogenous connectivity as well as condition-382 specific modulation between each respective region with VMPFC. Note that we focused our 383 DCM analysis on the posterior insula cluster only, as we were interested in a baseline frame 384 activation; including the anterior insula cluster, specific for generous choice within the loss 385 frame (see above), might have biased the results in favor of our hypotheses.

In total we defined 15 models (see Fig. 7) , grouped into two model families: A, which assumed of the log-evidence SF = -4.0786e+05, exceedance probability xp = 0.6508) (see supplemental Fig. S2b) , which assumed that the gain frame had an effect on the TPJ and its 394 connectivity with the VMPFC, while the loss frame had an effect on the VMPFC and its 395 connectivity with the insula (i.e., connectivity between regions is assumed to be bidirectional).

Concerning the driving inputs, we compared the average activity in TPJ in the gain frame 397 against 0, and the average activity in VMPFC in the loss frame against 0 (we checked, 398 beforehand, that no effect of repetition across runs was present; all ps > 0.18), but none of the 399 driving inputs was significantly different from 0 (all ps > 0.26; Table 1 ).

Next, when addressing the modulatory inputs, the only significant difference was found in the 401 loss frame for modulatory activity from the insula to VMPFC against the endogenous 402 connectivity from the insula to VMPFC (Mmodulatory = -0.2158 vs. Mendogenous = 0.04672, p = 0.016,

Bonferroni corrected), reflecting a significant modulation of endogenous connectivity by the 404 loss frame information (all other ps > 0.60; Table 1 ). In addition, the modulatory input was 405 negative, hinting towards an inhibitory influence of insula on VMPFC in the loss frame (as 406 before, there was no effect of repetition across runs in neither modulatory activity nor connectivity. Thick black lines represent unidirectional connectivity. Since endogenous connectivity is 412 always assumed between all three regions in all models, it is not represented here. Family A, which 413 assumed both condition-specific driving inputs and condition-specific modulatory inputs, includes 414 models 1 to 11. Family B, which assumed only condition-specific driving inputs, includes models 12 to 

To provide further support for our idea that the frame effect on social discounting was brought 428 about by a condition-specific neural activity pattern in the insula and VMPFC, and TPJ and 429 VMPFC, respectively, we ran mediation analyses on the relation between frame information, 430 social discounting behaviour, and neural activation in these regions. More specifically, frame 431 was entered as independent variable X (gain and loss), the hyperbolic k parameter (gain frame 432 and loss frame) was entered as dependent variable Y, and the neural activations were entered as mediators. We first focused on a model ( 

To clarify this finding even more, we built, next, a loss-specific model ( 

However, VMPFC seemed to play a crucial role as it mediated significantly the influence of 451 frame on k, and this mediation was significantly stronger than in combination with insula 

We provide behavioral and neural evidence for a simple nudge that aims at increasing 476 individuals' willingness to provide costly support to socially remote others. We adapted a social 477 discounting task where participants chose between a selfish optiona high gain to self and 

We indeed found that the anterior insula was significantly more activated when participants 514 made generous choices in the loss frame, relative to the gain frame, and our analysis further 515 showed that the degree of insula activation across trials correlated with the individual 516 propensity to make generous choices in the loss, but not the gain frame. Extending these 517 findings, we found that also the posterior part of the insula seemed to be involved in these 518 processes, specifically supporting the representation of the loss frame information even before 519 the decision was made (see also Droutman et al. 2015) . Building upon this evidence, we 520 further explored how both activation clusters mediated frame-specific social discounting 521 behavior. We propose and provide empirical support for a network model that predicts that the 522 frame effect on social discounting was associated with a frame-dependent neural connectivity 523 pattern between insula and VMPFC in the loss frame, and TPJ and VMPFC in the gain frame. 

Our analyses revealed two separate clusters within insula; while a more posterior cluster was 536 activated in response to general loss frame information, the more anterior cluster was specific 537 to generous choices in the loss frame. This topographic dissociation within insula is consistent 538 with previous findings suggesting a regional gradient in representing the level of abstraction 

Our findings expand on previous evidence that preventing harm to others is a great motivator away from selfish desires if these caused harm to others, including strangers. Our results normal vision. As reimbursement, they were paid €20 as participation fee, plus earnings from Social discounting task).

The study was conducted according to the Declaration of Helsinki and it was approved by the 

In the loss frame (Fig. 1b) , participants were informed, after the social distance presentation, 

In addition to the framing (gain frame, loss frame) and the social distance levels of the other

(1, 5, 10, 20, 50, 100), in each condition, we manipulated the magnitude of the own-reward 

Studies 1-3. All participants performed a social discounting task and, at the end, they participants were strongly encouraged to think as if they were making decisions for real. In 726 studies 1 and 2, participants were instructed about the social discounting task, and then, after 

To estimate V, we titrated the selfish amount to determine, at each social distance, the point 766 at which the subject was indifferent between the selfish and generous options (i.e., ( 1 * ß) + ( 2 * ß)

given the subjective values v (based on the current selfish amount and social distance) of the sequence; 208 sagittal images; voxel size = 0.8 × 0.8 × 0.8 mm; 0.80 mm slice thickness).

Head movements were minimized by the use of foam pads and scanner noise was reduced 807 with earplugs. When necessary, vision was corrected-to-normal via fMRI compatible goggles.

The social discounting task was programmed via an in-house software and presented via a 

Based on our univariate results, we constructed bilinear models (Fig. 7) where the 924 endogenous connectivity across the three regions was always assumed. We specified models 925 with nodes reciprocally connected, where the gain and loss frame were allowed to modulate 926 all connections (Li et al., 2015) . The resulting 15 models were grouped in two families: A and 927 B. In family A, both condition-specific driving inputs and condition-specific modulatory inputs 928 were assumed. In family B, only condition-specific driving inputs were assumed.

Family A included eleven models (Fig. 7) . In model 1, we assumed that the gain frame 930 condition had direct inputs on VMPFC and TPJ, and a modulatory input on their connections; 931 the loss frame condition had direct inputs on VMPFC and insula, and a modulatory input on 932 their connections. In model 2, the gain frame had a driving input on VMPFC, and a modulatory 933 input on its connectivity with TPJ; the loss frame had a driving input on VMPFC and a 934 modulatory input on its connectivity with the insula. In model 3, the gain frame had a driving 935 input on TPJ and a modulatory input on its connectivity with VMPFC; the loss frame had a 936 driving input on the insula and a modulatory input on its connectivity with VMPFC. In model 4, 937 the gain frame had a driving input on VMPFC and a modulatory input on its connectivity with 938 TPJ; the loss frame had driving input on the insula and a modulatory input on its connectivity 939 with VMPFC. In model 5, the gain frame had a driving input on TPJ and a modulatory input on 940 its connectivity with VMPFC; the loss frame had a driving input on VMPFC and a modulatory 941 input on its connectivity with the insula. Therefore, connectivity between regions in model 1 to on VMPFC and a modulatory input on its connectivity to TPJ; the loss frame had a driving 944 input on VMPFC and a modulatory input on its connectivity to the insula. In model 7, the gain 945 frame had a driving input on VMPFC and a modulatory input on the connectivity from TPJ to 946 VMPFC; the loss frame had a driving input on VMPFC and a modulatory input on the 947 connectivity from the insula to VMPFC. In model 8, the gain frame had a driving input on TPJ 948 and a modulatory input on its connectivity to VMPFC; the loss frame had a driving input on the 949 insula and a modulatory input on its connectivity to VMPFC. In model 9, the gain frame had a 950 driving input on TPJ and a modulatory input on the connectivity from VMPFC to TPJ; the loss 951 frame had a driving input on the insula and a modulatory input on the connectivity from VMPFC 952 to the insula. In model 10, the gain frame had a driving input on TPJ and a modulatory input 953 on the connectivity from VMPFC to TPJ; the loss frame had a driving input on VMPFC and a 954 modulatory input on its connectivity to the insula. In model 11, the gain frame had a driving 955 input on TPJ and a modulatory input on its connectivity to VMPFC; the loss frame had a driving 956 input on VMPFC and a modulatory input on the connectivity from the insula to VMPFC.

Family B included four models (Fig. 7) . In model 12, the gain frame had driving input on 

In model 1 (Fig. 8) , neural activations across both frames were treated as parallel mediators 985 and included the posterior insula (GLM2) and the anterior insula (GLM1), TPJ (GLM1), and 986 VMPFC (GLM1 and GLM2) (model template 4, 83). In model 2 (Fig. 8) 

(supplemental Table S4 ). Additionally, selfish amount magnitude, included as trial-by-trial 1474 Table S4 . Activations at selfish choice onset (button press), pooled across loss and gain frame.

Generosity and livelihoods: experimental DCM for fMRI