Cooperation across multiple game theoretical paradigms is increased by fear more than anger in selfish individuals

Cooperative decisions are well predicted by stable individual differences in social values but it remains unclear how they may be modulated by emotions such as fear and anger. Moving beyond specific decision paradigms, we used a suite of economic games and investigated how experimental inductions of fear or anger affect latent factors of decision making in individuals with selfish or prosocial value orientations. We found that, relative to experimentally induced anger, induced fear elicited higher scores on a cooperation factor, and that this effect was entirely driven by selfish participants. In fact, induced fear brought selfish individuals to cooperate similarly to prosocial individuals, possibly as a (selfish) mean to seek protection in others. These results suggest that two basic threat-related emotions, fear and anger, differentially affect a generalized form of cooperation and that this effect is buffered by prosocial value orientation.


Results
Induction validation. To induce fear, one group of participants took part in an anticipatory version of the Trier Social Stress Task 45 , involving a simulated job interview. Another group of participants took part in an Anger induction (inductions will henceforth be capitalized to distinguish these from other occurrences of the emotion terms), in which they received negative and unfair feedback on a short essay they wrote 67 . In a Control induction, a third group of participants were requested to read a passage of text. These inductions were matched in terms of duration and sequence of events ( Fig. 1A) (see "Inductions" in the Methods section for details). To validate these inductions, before and after being introduced to their respective activities, participants rated how well a number of words described their current mood, emotions and motives (see "Induction validation" in the Methods section for details). Inter-mixed amongst a number of control emotions and motives were words related to fear and anger. Changes in ratings were used as dependent variables to validate the inductions.
A linear mixed effect model on change scores (i.e., difference between post-induction and pre-induction ratings) revealed a significant interaction between the induction and the motive/affective state (F (16,1376) = 15.007, p < 0.001), suggesting that the latter were differentially affected by the inductions (Fig. 1B). Contrasts within the model suggested that the Fear and Anger inductions succeeded in differentially eliciting the target motives: increases in anger were higher in the Anger induction than in the Fear induction ("Anger- www.nature.com/scientificreports/ in happiness and sadness. To do so, we used the anger and fear change scores as dependent variables, and ran two linear regression models on each. As independent variable, we used the induction factor and the changes in happiness or sadness ratings. These models suggested that, even though happiness and sadness significantly contributed to the increases in anger and fear (all p s < 0.001, except for the impact of happiness on anger: p = 0.084), the inductions continued to reliably predict changes in the target motives (all p s < 0.001), even controlling for the non-target changes in happiness and sadness. This suggests that changes in fear and anger were far from fully explained by changes in happiness and sadness. Nonetheless, in addition to these control models, we further controlled for these unanticipated effects of our inductions by adding happiness and sadness change scores as covariates when modeling the economic decisions (see "Model 2" in the Supplementary Material, Supplementary  Tables ST4 and ST5).
Two factors of economic behavior. While participants waited for the induction activities to take place, they took part in an allegedly separate study on economic decision making, involving real monetary incentives, and consisting in a suite of game theoretic paradigms (see "Game theoretic paradigms" in the Methods section for details). In contrast, social value orientation measures (SVO) were requested from participants 2 weeks after the inductions to avoid any spill-over effects between the two (see "Social value orientation" in the Methods section for details). Finally, an 'experimental demand' questionnaire was conducted after the game theoretic paradigms to probe participants' awareness of any relation between these and the inductions (see Supplementary Material SM1, "Experimental demand questionnaire"). A Spearman's correlation matrix (Fig. 2) suggested that two groups of economic behaviors were inter-correlated. Specifically, 1st and 2nd movers in the trust game-represented by average (1st mover) transfer rates (i.e., "trust" by 1st movers) and average returns (i.e., "trustworthiness" by 2nd movers), average charitable donations, transfers in the dictator game, proposals as 1st movers in the ultimatum game, contributions to public goods, restraint in common resource dilemmas and helping in the Zurich prosocial game, all positively correlated. On the other hand, average punishment rates in the 2nd and 3rd party punishment games, and frequency of rejections in the ultimatum and impunity games, all positively correlated with one another.
To formally test these groupings in a factor analysis, we first obtained a number of indexes indicating how many components to retain (see "Statistical analyses" in the Methods section for details). All tested indexes (parallel analysis, optimal coordinates, acceleration factor, very simple structure and Velicer's MAP criterion) recommended retaining 2 factors. Moreover, since oblique (i.e., oblimin) rotation revealed only a very weak This two factor solution largely corroborated the groupings of economic variables informally suggested by the correlation matrix (Table 1 and Fig. 3): amounts entrusted and returned (i.e., "trustworthiness"), charitable donations, offer sizes in the ultimatum game, transfer size in the dictator game, contributions to public goods, restraint in common resource dilemmas and helping in the Zurich prosocial game, all loaded on the first factor (all loadings > 0.35). Following previous studies 55, 56 , we labeled this a "Cooperation factor". On the other hand, amounts spent to punish others in 2nd and 3rd party punishment games, frequency of rejections in ultimatum and impunity games loaded on the second factor (all loadings > 0.45). Since the top loading variables on this factor were the 2nd and 3rd party punishment games, we called this a "Punishment factor". Most variables loaded uniquely on the respective factors (mean item complexity = 1.1). Only the rDOC and the DG had above higher complexity (1.5 and 1.3, respectively), loading negatively and positively on the punishment factor, respectively. Böckler et al. 54 describe two similar factors as an "altruistically motivated prosocial behavior" factor and a "normmotivated prosocial behavior", respectively.
The impact of induced fear and anger on latent constructs of cooperation and punishment. The factor analysis additionally enabled to obtain one pair of scores for each participant. These can also be thought of as a pair of coordinates, determining participants' position on a "cooperation x punishment space" (Fig. 3). These two scores were submitted to regression analyses. Our primary model of interest (see "Model 1" in "Statistical analyses") on the cooperation scores revealed a highly significant impact of SVO on cooperation (F ( Table 1. Factor analysis of 12 economic variables: 2-factor solution. Standardized loadings (pattern matrix) based upon correlation matrix, communality ("Com") and complexity ("Comp") of each variable. Two factors, labeled "Cooperation" and "Punishment", captured 31% of the total variance.

Discussion
Human cooperative behaviors are modulated by stable inter-individual differences in prosocial or selfish value orientations, yet it remains unclear how these stable values interact with induced or incidental perceptions of threat. To address this, we experimentally induced either anger or fear in individuals with different social value orientations and had them take part in a suite of incentivized decision paradigms that fall along one of two latent factors: cooperative and punishment based decisions 4,27,54-56 . Our results show that induced Fear increases scores on a latent cooperation factor, relative to induced Anger. We also find that this effect is entirely driven by individuals with "selfish" social value orientations. These results extend previous findings in several ways. In non social decisions, induced fear and anger have been respectively associated with avoidance and approach of risk 18,19,[21][22][23][24][25][26] . Our results extend this by showing that, in the context of social decision making related to cooperation, induced fear is more likely to increase cooperation, relative to induced anger. In addition, given that fear and anger both involve negative valence and high arousal, our findings suggest that the impact of these induced emotions on decision making are unlikely to be entirely explained by mood or arousal alone 19 , nor by "mood reparation" 48 . Rather we suggest that these findings are consistent with motive-based approaches to fear and anger, which link fear and anger to particular motives, such as a defensive or aggressive motives, respectively 15,69 . Under this view, when social contexts are involved, induced fear may lead to social approach oriented behaviors as a protection mechanism, to seek support and acquaintance in the face of danger 38 ; while anger can lead subjects to avoid cooperative gestures, as a means to antagonize others 28 .
The fact that this differential effect of induced Fear and Anger is only observed in selfish individuals aligns with the notion that strong cooperative values can buffer the impact of certain motives and emotions 59 : while selfish individuals could behave cooperatively especially when it suits their need for self-protection 44,48 , such a self-serving motive is incoherent with prosocial value orientations. This is also in line with evidence suggesting that fear-induced cooperation is unlikely to stem from genuine care for others. For instance, perceived threats have been shown to increase cooperation with one's peers or ingroup 70,71 and stress has been shown to increase www.nature.com/scientificreports/ cooperation only towards closer others 42 . This instrumental perspective of the "tend and befriend" hypothesis 38 offers a plausible explanation of why this cooperative response to fear is only observed in selfish individuals. Surprisingly, we do not observe a main effect of induced anger on decisions to punish, which has instead been observed in previous studies [31][32][33][34] . With hindsight, we speculate that this can be due to an incorrect selection of a moderating individual difference variable for the punishment domain. In fact, while there was a clear candidate for the value orientation that would predict cooperative behavior, the same was not true for punishing behaviors. Future studies might address whether "social dominance orientation" 72 , or a moderating variable that has been shown to predict punishments in game theoretical paradigms (such as the 'assertiveness scale' 73 ), may be able to better detect the interaction of values and incidental emotions on punishments.
Finally, by taking a factor analytic approach our study is the first to our knowledge to document a link between induced motivational states, social values, and a "domain general" cooperation factor 55 . This approach does not intend to deny important differences between individual decision paradigms, some of which might very well influence how specific decisions are affected by motives and emotions. Rather, we suggest that this approach could provide a relatively unbiased illustration of the effects of induced Fear and Anger on cooperative decision making, by describing how these threat-related emotions affect a source of behavioral variance that is common to many if not all of them 4,54-56 . This factor analytic approach might thus alleviate problems related the reliability and ecological validity of game-theoretical paradigms 53 . Ultimately, it may also contribute to the ongoing integration of economics and psychology 9,74,75 , by helping economists formalize the impact of emotions and motives on economic decision making 76 .

Conclusions
Plausibly, no psychologist would endorse one of the basic assumptions of neoclassic economic theory: that decisions are driven only by stable context-insensitive preferences 9,77 . However, to integrate psychological and economic frameworks, psychologists should provide empirical evidence as to how important motives and emotions such as fear and anger systematically affect economic decision making. Our results suggest that in individuals with selfish value orientations induced Fear is more likely to increase cooperative behaviors, relative to induced Anger. The finding that selfish individuals drive this differential effect suggests that prosocial values can buffer the impact of certain emotions on cooperation: while selfish individuals may especially increase their cooperative behavior in the face of fear, such instrumental cooperation may conflict with the value orientations of prosocial individuals. Taken together, these findings shed light on the importance of both context-sensitivity and pre-existing differences in prosocial value orientations in determining cooperative behavior. Finally, by capitalizing on a source of variance that is common to many decision contexts and on inter-individual differences, our results highlight how fear-inducing contexts (e.g., cultures or climates of fear) can influence a generalized form of cooperative behavior, and sheds light on which individuals might be more susceptible to this.

Methods
Participants. 175 participants (82 males, 93 females, mean age = 27.1, SD = 4.8) were assigned either to a "Fear" (N = 56, 25 males, 31 females, mean age = 26.5, SD = 4.4), an "Anger" (N = 56, 26 males, 30 females, mean age = 26.6, SD = 4.1) or a "Control" group (N = 63, 31 males, 32 females, mean age = 27.8, SD = 5.7). A separate group of participants was assigned to two other motive-inductions not relevant to this study, which have been published elsewhere together with the same Control group 56 . Participants were recruited through the Max Planck's Institute participant database. All studies were advertised, via email, to all eligible participants in the database, namely, participants between the age of 18 and 65, with no history of cognitive, psychiatric or neurological disorder. Participants registered to the studies on a first-come first-served basis. Age and gender were similarly distributed across each group (all pairwise comparisons, p s > 0.1). Participants provided informed consent for the treatment of their anonymized data and all methods were carried out in accordance with relevant guidelines and regulations. Assessments were approved by the Research Ethics Committee (Agreement Number 090-15-09032015) of the University of Leipzig, Germany. Data will be made available upon reasonable request.
Inductions. Group sessions took place in a computer room with shielded computer cubicles. In all inductions, participants first provided baseline ratings on a number of emotion and motive-related items (see "Induction validation"). Then, they were informed about one of three activities. Participants in the Fear group took part in the anticipatory Trier Social Stress Task (henceforth, "aTSST") 45,78 , in which they were informed about a simulated job interview requiring them to take part in a series of (stressful) tasks in front of two anonymous interviewers. We chose the anticipatory variant of the TSST because while the standard TSST is known to elicit both fear and anger 79 , the aTSST has been suggested to predominantly enhance anxiety rather than hostility and aggression 78 . To induce Anger, we adapted the "negative feedback" procedure 67,80 in which participants receive negative (and unfair) feedback on a short personal essay they wrote and anticipated providing feedback to the reviewer's essay in turn. In the Control condition, participants only anticipated reading a passage of a text. To increase the salience of these activities, participants were accompanied to a different room, one by one. Here, in the Fear induction, they found the two aforementioned interviewers (one male, one female), wearing lab-coats and sitting behind a desk. These interviewers asked participants a set of preliminary questions in a detached fashion (e.g., what job they would like to interview for, and why they thought they would be good candidates for that job). In the Anger induction, participants were shown a single-blind mirror, where they later would have the opportunity to provide feedback to their (unfair) reviewer, via microphone, on his/her essay. In the Control induction, subjects were shown the recording room where they later would be recorded whilst reading the text passage and gave a brief sound check. After returning to the computer room, subjects provided ratings on the same emotion and motive-related items rated previously. A different experimenter then told participants www.nature.com/scientificreports/ that, while they waited for the activities to be prepared, they would take part in an allegedly different study on economic decision making (see "Game theoretic paradigms", below). To maintain the inductions salient, half way through the decision making study, participants were asked to take notes for 5 min, to prepare for the activities. After completing the decision making study, participants took part in a written questionnaire probing for awareness of any connection between the two studies (see Supplementary Material SM1, "Experimental demand questionnaire"). Finally, participants were fully debriefed and paid for their participation and for one of their decisions. The whole experimental session was self-paced and lasted 1 h and 30 min on average.

Induction validation.
Before and after being introduced to the induction-specific activities (see "Inductions"), participants provided baseline ("pre-induction") and test ("post-induction") ratings indicating how well a list of fear and anger-related words described their motivation or feelings (with visual analogue scales ranging from − 350 to 350). The fear items were (here translated from German) "apprehensive", "afraid", "timid", "nervous", "panic-stricken", "overcautious", "frightened", "reserved", while the anger items were "aggressive", "angry", "offended", "irritable", "argumentative", "tempestuous", "spirited". These items were chosen because they bestdiscriminate the target constructs from related but distinct constructs 81 . Previous work using emotional inductions prevalently controlled for constructs of interest only, yet it appears plausible that, an Anger induction may concurrently increase feelings of power in some individuals, or that a Fear induction may also increase feelings of achievement, given that it involves a simulated job interview. To address this, in addition to the emotions of interest (i.e., fear and anger), the questionnaire also probed five motive-related measures: achievement, affiliation, care, power and consumption; and two affect-related measures: happiness (i.e., positive affect) and sadness (negative affect) (see Supplementary Material SM2 for the list of all items). Overall, the questionnaire consisted of 63 items composing 9 measures (7-motive related and 2 affect related constructs). The order of all items was fully randomized for each participant, who rated 7 items per page. To analyze this data, we first subtracted the pre-induction ratings of each item from the corresponding post-induction ratings and then averaged over items pertaining to the same construct, thus yielding 9 "change scores". Finally, to validate the inductions, we compared these difference scores between the inductions (see "Statistical analyses").

Game theoretic paradigms.
To measure cooperation and punishment we had subjects take part in a suite of paradigms that have been found to factor together in previous studies [54][55][56] . Specifically, as candidate contributors to a "cooperation factor", we had participants take part in a dictator game (in which they decided how much money, if any, to transfer to a passive recipient), a trust game (in which, as 1st movers, participants decided how much to entrust to a second mover, and as 2nd movers decided how much to return to the 1st movers), a charitable donations game (in which participants decided how much to donate to various charities), a public good game (in which they decided how much money to contribute to a public good), a common resource dilemma (in which they decided how much to take from a common resource) and the "Zurich prosocial game" (in which participants decided how much to help their counterparts in a virtual maze). As for the punishmentrelated games, we adopted the 2nd and 3rd party punishment game (in which participants observed how much another player transferred to themselves or a third party-2nd and 3rd mover variants, respectively-and, on the basis of this, decided how much money, if any, to invest to "punish" them, that is, to decrease their payoff), an ultimatum game (in which, as 1st movers, participants made a proposal on how to split a monetary prize to a responder, knowing that responders would then have two options: if they accepted the monetary prize was split as proposed, if they rejected, both players received nothing), and an impunity game (which is identical to the ultimatum game with the exception that, if second movers rejected an offer, first movers still retain what they proposed to keep for themselves). Following our previous study 56 , we first obtained one measure for each paradigm (averaging over measures in the case of multiple trials). We then subjected the resulting 12 measures to factor analysis (see "Statistical analyses"). Finally, to control for the potential impact of our inductions on social decisions net of any effect they may have on non-social decisions, we measured risk attitudes by means of a lottery task. This task was taken from Dawans and colleagues 39,45 , and involved a series of binary decisions in which participants chose between two lotteries with similar expected value but different levels of risk. The frequency of choices of the riskier lottery was used as a measure of participants' non-social risk attitudes. All economic games were divided in two blocks to avoid potential spill-over effects between superficially similar games (such as 2nd and 3rd party punishment, see Supplementary Material SM3 for full details on block composition). The order of games within a block was fully randomized and the order of blocks was counterbalanced between participants. Participants were informed that they would be paid for one randomly determined decision, at the end of the experiment. Full details on each of the decision paradigms are available in the Supplementary Material (Supplementary Tables ST1 and ST2). Instructions are available upon request.

Social value orientation.
To measure inter-individual differences in social values we used the social value orientation task ("SVO") 2,3 . In order to reduce the possibility that the inductions could affect the SVO scores, these were measured on a different day. Specifically, two weeks after the participants came to the lab to take part in the economic decisions, they were sent an email linking them to the SVO questionnaire (to be done online). We used the SVO "slider measure", as this has been suggested to be more reliable than previous measures 82 and we focused on the 6 primary items because the secondary items are mostly required to distinguish between finer types of prosocial orientation, which were outside of the scope of this study. Each of these six primary items require subjects to choose between nine different point allocations to themselves and anonymous others (e.g., between option A: {100 for self and 50 for other} vs. option B: {85 for self, 85 for others}, etc.). Following previous research 83 , we divided participants into "proself ", also called "selfish" participants (a classification which combines subjects displaying individualists and competitive values), and "prosocial" participants (combining par- www.nature.com/scientificreports/ ticipants with prosocial and altruistic values). These aggregations are typically made because of the relatively low numbers of altruists and competitors that are observed 5 . All participants had accepted to take part in this additional online component of the experiment but 20 participants (three in the Anger induction, three in the Fear induction, 14 in the Control induction) never got back to our invitation emails. Consequently, the SVO scores of these participants, but not their economic decisions, are missing. The SVO scores did not differ between any of the groups (all p s > 0.17). In the Anger, Control, and Fear conditions, respectively, we identified n = 17, 23 and 16 proself participants and n = 36, 26 and 37 prosocial participants. Finally, our results also held when modelling SVO as a continuous predictor (Supplementary Table ST6and Supplementary Fig. SF1) 84 .

Statistical analysis.
For induction validation, we used a mixed effect model predicting change scores in self reported ratings based on the induction (with levels: "Fear", "Anger", "Control") the motivational/affective state (9 levels) and their interaction. These factors were modeled as fixed effects, while participant IDs were used as random intercept terms, to account for the fact that observations were clustered at the subject level. Planned contrasts within this model focused on the emotions of interest, namely, whether changes in fear were higher in the Fear induction relative to the other inductions, and whether changes in anger differed between the Anger induction and the other inductions. Additional exploratory contrasts investigated the remaining 7 change scores.
To investigate our main question of interest (i.e., whether induced Fear and Anger differentially affect latent factors of social decision making), we first aimed to obtain a reliable factor scores for the 12 decision environments, following the methods used in our previous work 56 . More specifically, we first investigated the optimal amount of factors to retain using a number of standard indices 68 , and then tested whether the resulting factor structure was stable across the inductions 85 . Factor analysis was performed with the "fa" function (in the "Psych" package) 86 , while the stability of the factor structure was performed with the "measurementInvariance" function (in the "semTools" package) 85 . Finally, we extracted participants' scores on each of the obtained factors, and investigated whether these differed between the inductions, using multiple regression models. We tested two models on each factor score. "Model 1" addresses our main hypotheses of interest, thus predicting participants' factor scores on the basis of the induction (three levels: Fear, Anger and Control), participants' SVO category (two levels: prosocial and proself), and their interaction. "Model 2" aimed to assess robustness of the results by additionally controlling for a number of potential confound variables, including gender-which has elsewhere been shown to have a potential impact on decisions to cooperate or punish 71,87 -as well as the (standardized) measure of subjects' risk attitudes as measured by the lottery task described above. We report the results of Model 1 in the results section and the results of Model 2 in the Supplementary Material. For all models, we report the results of the omnibus tests (as assessed by type III Anova) and further qualify these by means of contrasts, of which we report 95% confidence intervals p-values (as computed in the "lsmeans" package 88 ). P-values are Bonferroni corrected, unless otherwise noted. All analyses were carried out in R 89 . Stimuli were prepared and administered in Presentation (Neurobehavioral Systems, Inc.).