Fig 1.
Participants observed what they were led to believe were other participants’ resource allocation decisions. On each trial, the Experimenter gave a financial endowment (£1 to £15, step size = £1) to Player B. Experiment 1 (A): Player A, the allocator, could then take a portion of this money for themselves (10% to 100%, step size = 10%). In blocks 2 and 4 participants rated how they felt about the allocator’s decision. In all blocks, participants decided if and by how much to punish the allocator. In blocks 2 and 4 participants subsequently rated how they felt about their punishment decision. Experiment 2 (B): Player B was the allocator and could share a portion of their money with Player A (0% to 90%, step size = 10%). In blocks 1 and 3 participants rated how they felt about the allocator’s decision. In blocks 2 and 4 participants decided if and by how much to punish the allocator and then rated how they felt about their punishment decision.
Fig 2.
Feelings are negative when observing selfish behavior, positive when observing equal splits, and neutral when observing generosity.
Observers report negative feelings when observing allocators act selfishly and positive feeling when observing equal splits. Surprisingly, feelings were not positive (nor negative) when observing generosity. This is true both in Experiment 1 (A, left panel) and Experiment 2 (B, right panel). The grey dots represent the mean feelings rating per allocator type for each participant. Red diamonds represent the average of these means. The box plots show the distribution of the participants’ mean feelings about each allocator: boxes indicate 25–75% interquartile range, whiskers extend from the first and third quartiles to most extreme data point within 1.5 × interquartile range, and the median is shown as a horizontal line within this box. *** p < 0.001.
Fig 3.
Operationalizing selfishness and inequality.
Selfishness (blue line) is defined as the percentage of the endowment the allocator took/kept for themselves ranging from 0 (when the allocator is most generous, allocating nothing to themselves) to 100 (when the allocator is most selfish, allocating all to themselves). Inequality (orange line) is defined as the absolute difference between the percentages of the endowment each person is left with post allocation ranging from 0 (when the split was 50/50) to 100 (when one person receives all and the other none).
Fig 4.
(A) Model specification. Each row corresponds to a model, each column represents a parameter. Grey colour indicates that the parameter is included in the model. For example, the first row corresponds to model 1, which include selfishness only. Note that all the models also include a constant. (B, C) Model selection. Log delta BIC (difference between each model and the best model) is shown for (B) Experiment 1 & (C) Experiment 2. The green circle indicates the best model. Non-transformed ΔBIC between best model and second-best model > 50. (For additional information on model fit metric and ranking see Table A in S1 Text.) (D, E) Parameter recovery. Each row corresponds to a model and the columns represent the regressors. Coloured values correspond to the Pearson correlation r between the true parameters that generated the data and the estimated parameters in Experiment 1 (D) and 2 (E). The black colour is used when there is no parameter for this model. All the Pearson r are significant at p < 0.0001. (F, G) Model recovery analysis. The x-axis shows the model number which was used to simulate data and the y-axis the model number which was fit to the simulated data. The black color shows which model best fit the simulated data (compared to the second-best model using a ΔBIC > 30) for feelings simulated data in Experiment 1 (F) and Experiment 2 (G). The diagonal line indicates perfect model recovery. In other words, the model used to simulate the data was also the model that best fit that data. See methods for details.
Fig 5.
Modelling observers’ feelings as a function of observed selfishness and inequality.
Plotted is the winning model (model 26) fit at the group-level for Experiment 1 () and Experiment 2 (
). Participants’ feelings ratings were z-scored before model-fitting to standardized responses. The predicted feelings from the model (red line) is overlaid on the mean observed feelings over all participants (black dots). Error bars represent SEM.
Fig 6.
Participants punished selfish allocators more often (A & B) and more severely (C & D) than generous allocators and those who split resources equally in both Experiment 1 (A & C) and Experiment 2 (B & D). Surprisingly, participants also punished generous allocators more frequently than allocators who split the money equally in Experiment 2 (B & D) and punished them the same in Experiment 1 (A & C). Grey dots show the proportion of trials on which each participant punished each allocator (top panels) and the mean amount of punishment per allocator for each participant (bottom panels). Red diamonds represent the averages of these means. The box plots show the distribution of the participants’ mean punishment per allocator: boxes indicate 25–75% interquartile range, whiskers extend from the first and third quartiles to the most extreme data point within 1.5 × interquartile range, and the median is shown as a horizontal line within this box. * p < 0.05, *** p < 0.001.
Fig 7.
Model specification. (A) Model specification. Each row corresponds to a model, each column represents a parameter. Grey colour indicates that the parameter is included in the model. For example, the first row corresponds to model 1, which include selfishness only. Note that all the models also include a constant. (B, C) Model selection. Log delta BIC (difference between each model and the best model) is shown for (B) Experiment 1 & (C) Experiment 2. The green circle indicates the best model. Non-transformed ΔBIC between best model and second-best model > 50. (For additional information on model fit metric and ranking see Table B in S1 Text.) Parameter recovery. Each row corresponds to a model and the columns represent the regressors. Coloured values correspond to the Pearson correlation r between the true parameters that generated the data and the estimated parameters in Experiment 1 (D) and 2 (E). The black colour is used when there is no parameter for this model. All the Pearson r are significant at p < 0.0001. Model recovery analysis. The x-axis shows the model number which was used to simulate data and the y-axis the model number which was fit to the simulated data. The black color shows which model best fit the simulated data (compared to the second-best model using a ΔBIC > 30) for feelings simulated data in Experiment 1 (F) and 2 (G) The diagonal line indicates perfect model recovery. In other words, the model used to simulate the data was also the model that best fit that data. See methods for details.
Fig 8.
Observers’ punishment decisions reflect selfishness aversion and inequality aversion.
Plotted are the winning models fit at the group-level for Experiment 1 ( left panel) and Experiment 2 (
right panel). Participants’ punishment choices were z-scored before model-fitting to standardize responses. The predicted punishment from the model (purple line) is overlaid on the mean observed punishment over all participants (purple dots). Error bars represent SEM.
Fig 9.
Transformed standardized coefficients (i.e. punishment coefficients are reversed for comparison) revealed a greater effect of selfishness on punishment than feelings, with no difference in the effect of inequality.
In addition, reactions were more influenced by observing selfish taking than selfish giving, but vice versa when observing inequality. Furthermore, social values effects punishment more than feelings and selfishness effected responses more than inequality. Error bars correspond to the SEM.**p<0.01,****p<0.0001.