Neural Correlates of Gender Differences in Reputation Building

Gender differences in cooperative choices and their neural correlates were investigated in a situation where reputation represented a crucial issue. Males and females were involved in an economic exchange (trust game) where economic and reputational payoffs had to be balanced in order to increase personal welfare. At the behavioral level, females showed a stronger reaction to negative reputation judgments that led to higher cooperation than males, measured by back transfers in the game. The neuroanatomical counterpart of this gender difference was found within the reward network (engaged in producing expectations of positive results) and reputation-related brain networks, such as the self-control network (engaged in strategically resisting the temptation to defect) and the mentalizing network (engaged in thinking about how one is viewed by others), in which the dorsolateral prefrontal cortex (DLPFC) and the medial (M)PFC respectively play a crucial role. Furthermore, both DLPFC and MPFC activity correlated with the amount of back transfer, as well as with the personality dimensions assessed with the Big-Five Questionnaire (BFQ-2). Males, according to their greater DLPFC recruitment and their higher level of the BFQ-2 subscale of Dominance, were more focused on implementing a profit-maximizing strategy, pursuing this target irrespectively of others' judgments. On the contrary, females, according to their greater MPFC activity and their lower level of Dominance, were more focused on the reputation per se and not on the strategic component of reputation building. These findings shed light on the sexual dimorphism related to cooperative behavior and its neural correlates.

, where nvoxL is the number of active voxels in a cluster in the left cerebral hemisphere, and nvoxR is the number of active voxels in the homologous region of the right hemisphere.

Preliminary fMRI analysis and results
In order to investigate the difference between Positive and Negative evaluation in the Reaction phase, we performed a preliminary GLM analysis using a more complex model than that presented in the Main Text (MT), which also included the two levels (Negative or Positive) of the Reaction phase as regressors. However, the contrast between Negative and Positive reputation did not lead to any significant difference in brain activations. This is consistent with Liu et al. (2011) meta-analysis on reward valence, suggesting that some components of the reward network are commonly activated by both positive and negative rewards across various stages of reward processing (e.g., anticipation, outcome and evaluation).
We were unable to perform a GLM including the three behavioral levels of back transfer (nothing; the same amount sent by A; an amount that equalize payoff between A and B) as regressors, because only 7 out of 16 subjects used all the three possible choices during the game. For the other 9 subjects one level, depending on the subject, was missing. Thus, in order to control the behavior effects in our data, we performed the ANCOVA described in the MT.

Path Analysis with Structural Equation Modeling
We investigated whether gender has a significant direct impact on behavior by means of a path analysis conducted with Structural Equation Modeling (SEM).
In particular, we extracted a dataset composed by trial by trial beta values of activation in 8 brain areas. Selected areas are those that, in the ANOVA analysis have been found significantly covariant either with Gender (Caudato and Insuala) or with the interacted variable between Reputation and Gender (precuneus, fusiform gyrus, DLPFC and DMPFC) in the Choice phase and with Gender (fusiform gyrus, VLPFC) in the Reaction phase. We also included behavioral choices and information about subjects' gender, treatments and runs of the experiment and we obtained an unbalanced dataset with a total of 690 observations for the 16 subjects.
In SEM we consider four exogenous variables: Gender (Male or Female), Reputation (Reputation or No Reputation), the interacted variable between Gender and Reputation and the variable indicating the run of the experiment. Endogenous variables are activation values in the 8 brain areas introduced above and one behavioral variable that represents the choice made by subjects and it is measured as the rate between the units given back to investors and the total endowment available.
We developed a SEM model starting from the following assumptions: 1) all brain areas have a direct effect on behavior, as suggested by the data analysis presented throughout the paper; 2) brain areas can be correlated and this can be modeled only in terms of covariance of error terms (i.e., in the model there cannot be any directed causal relationship between brain areas); 3) any exogenous variable can directly predict any endogenous variable.
Following these hypotheses, a SEM model is developed by means of an iterative procedure lead by a criterion of model fitness. The procedure starts from a random configuration of relationships between variables, then causal links and covariances among error terms of endogenous variables are removed if not significant (p > 0.05) and added if their modification index points out a significant improvement (p < 0.05) of the model in terms of reduction in its χ 2 .
Following the assumptions above, direct effects between brain areas and the behavioral variable are kept constant throughout the procedure. Both covariances between error terms of endogenous variables and causal links directed from exogenous variables to endogenous ones are introduced and kept in the model only if they significantly improve the capability of the model to fit the data. The procedure stops when modifications improving the model further and significantly are not available. The resulting SEM model does not significantly fail in reproducing the covariance matrix for the 13 variables considered, in fact its χ 2 is small relative to the degrees of freedom and not significant (p > 0.05). Moreover, the probability of the model to have a root mean squared error of approximation less than or equal to 0.05 is 0.998 and its Comparative Fit Index is 0.984. 3) Reputation has a direct effect on PCN, Caudato and VLPFC; 4) Gender has a direct effect on DLPFC, Caudato, Insula, Fusiform (reaction phase) and VLPFC; 5) The interacted variable between Gender and Reputation has a direct effect on Fusiform (choice phase), DLPFC, DMPFC and Caudato; 6) All brain areas have a significant (p < 0.05) direct effect on the behavioral variable, except Caudato and Fusiform (reaction phase); 7) The Gender and Reputation variables have a significant indirect effect on the behavioral variable (respectively, p = 0.005 and p = 0.000).
To sum up, the SEM model showed that gender does not have a significant direct effect on behavior.
Nevertheless, gender does have a significant indirect effect on behavior, that is to say that gender predicts behavior only if mediated by brain areas.
Also different models were developed and all of them confirmed the results of the model described above. Results are confirmed when the behavioral variable is measured either as the absolute amount of units given back to investors or as its difference with the average individual amount of restitution in the baseline treatment. Similarly, again, results are also confirmed when relaxing the first assumption described above. In fact, we developed a further model where direct effects spanning from brain areas to behavior are evolved according to model fitness instead of imposing a constant structure. All these models do not contain a direct effect of gender on behavior.
In all models, estimation of parameters has been conducted with a maximum likelihood algorithm considering missing values (the behavioral variable is missing when subjects did not make a choice in time or pushed the wrong button during the experiment). In all models, the variance-covariance matrix of the estimates has been estimated by means of the generalized Huber-White-Sandwich estimator considering each subject as a single group (i.e., cluster-robust standard errors).
Further details about SEM models are available from authors on request.

fMRI and PET Meta-analysis
We generate three meta-analyses of interest using the BrainMap database, the Sleuth 2.0.3 and  See below the list of references included in the three meta-analyses and see results in Tab. S2 and

Game dynamics
The game dynamic showed a systematic increase of back transfers in the Reputation treatment rounds (Fig. S3). In both runs, the reputation conditions led to similar averages (

Behavioral effects of the received evaluations
Overall subjects received a total of 288 evaluations, with a majority of positive ones (163 positive vs. These results are confirmed by a fixed effects model using the first order differences in back transfer as dependent and the evaluations received (in comparison with the no reputation treatment) as independents, controlled by the amount sent by A players. All coefficients are highly significant except the one related to the condition where subjects did not yet receive an evaluation, which is significant only at the 10% level. Consistent with the analysis above, subjects increased their back transfer following a negative evaluation, but decreased them following a positive one (Tab. S3).

Payoffs
Average payoffs were slightly lower in the reputation rounds (25.4 ± 0.6 MU/round) than in the baseline ones (29.9 ± 0.7 MU/round) due to the fact that B players who transferred back a higher        Table 2 in the MT).