Favoritism and cooperation

All experimental instructions and questionnaires are provided in Online Appendix A.

The participants who were not selected for the high-stake group played a similar game in Part 3, but were endowed with only half the amount of money that the members of the high-stake group received. The members of the high-stake group never interacted with those who were not selected. For the results presented in the main text, data were processed and analyzed only for the high-stake group for two reasons: (1) the experiment was designed specifically to study the behavior of those selected by a process involving favoritism and (2) the color composition of the non-selected groups varied (by design) between the favoritism and the control treatment. For a brief analysis and discussion of the data on the non-selected participants, please see Online Appendix B.

Information was collected about the participants’ sex, age and race. No significant differences in demographic characteristics between the treatment and the control groups were observed, with all results reported here robust to including demographic controls.

All standard errors reported in Sects. 3 and 4 are clustered at the individual level.

This p value, and all other p values reported here, come from two-sided Wald tests with standard errors (SEs) clustered at the individual level. All reported results are robust to exchanging the Wald test for a non-parametric Mann–Whitney test.

The impression that Fig. 1 gives is that the favoritism treatment primarily affects participants’ propensities to cooperate or defect fully. We can investigate that possibility further by typing participants, and doing so reveals that the share of participants who always contribute all of their endowment to the public good increases from 2.5% in the favoritism treatment to 20% in the control treatment (the difference is statistically significant on a t test of proportions: p = 0.013). The share that never contribute anything to the public good falls from 27.5% in the favoritism treatment to 12.5% in the control treatment (the differences is, however, only marginally significant: p = 0.096). Excluding participants who either contribute everything or nothing in all games reveals that the remaining participants contribute 43.2% (SE = 3.5) of their endowment to the public good in the favoritism treatment and 51.3% (SE = 4.0) in the control treatment (the difference is not significant: p = 0.129).

A further discussion about a possible link between math task performance and PG game behavior is provided in Online Appendix D.

The just-reported finding is in line with previous research on in-group favoritism, assuming that participants regard players with the same color as themselves as the ingroup, whereas other participants belong to the outgroup (cf. Tajfel 1970; Tajfel et al. 1971; Bernard et al. 2006; Goette et al. 2006; Chen and Li 2009). However, the same ingroup effect is present and of similar size in both the favoritism and the control treatments. Another way is available to investigate if the general tendency of participants to contribute less when playing with people who were assigned the other color affected the difference between the favoritism and the control treatment. To do that, we utilize the fact that the order in which the games were presented varied such that some participants had game number 1 against a person with the same color as herself and others did not. (Participants were presented with, and made decisions in, all seven games simultaneously, but the games were numbered from 1 to 7 and most people filled them out in numerical order.) We can then check for order effects, i.e., if participants whose first game involved someone of the same color behaved differently against all players (e.g., anchored at a higher cooperation level) than those whose first game was against a player of the other color. The analysis in Online Appendix C shows that no such order effects are found in the data.

The following question was asked in the post-experimental questionnaire: “since [argument] it was the right thing to do to select four purple players and four orange players into the high-stake group” (that question was asked only of the participants in the justification treatments, argument as described in main text, see instructions in Online Appendix A for exact formulation). On a 1–10 scale where 1 = strongly disagree and 10 = strongly agree, the efficiency and fairness justifications received a score of 6.5, SE = 0.6 and 7.9, SE = 0.3, respectively. The difference is statistically significant (p < 0.05). Moreover, participants who saw the fairness justification found the selection procedure to be fair more often than those who saw the efficiency justification (87.5%, SE = 6.9 and 50.0%, SE = 10.4, respectively). That difference is statistically significant, p < 0.01.

Note that the comparison implies, in the point estimates, less cooperation in the control treatment with random selection than in the control treatment with performance selection. That difference is not statistically significant; see also Online Appendix D, part 2.