Reputations for treatment of outgroup members can prevent the emergence of political segregation in cooperative networks

Abstract

Reputation systems promote cooperation and tie formation in social networks. But how reputations affect cooperation and the evolution of networks is less clear when societies are characterized by fundamental, identity-based, social divisions like those centered on politics in the contemporary U.S. Using a large web-based experiment with participants (N = 1073) embedded in networks where each tie represents the opportunity to play a dyadic iterated prisoners' dilemma, we investigate how cooperation and network segregation varies with whether and how reputation systems track behavior toward members of the opposing political party (outgroup members). As predicted, when participants know others' political affiliation, early cooperation patterns show ingroup favoritism. As a result, networks become segregated based on politics. However, such ingroup favoritism and network-level political segregation is reduced in conditions in which participants know how others behave towards participants from both their own party and participants from the other party. These findings have implications for our understanding of reputation systems in polarized contexts.

Fig. 1. Average cooperation.

Cooperation is measured by number of monetary units sent in each round. All rounds for each experimental condition are given in (A). The sample size of A is 67,774 instances of cooperation, nested in rounds, nested in participants. Marginal cooperation (number of monetary units sent) for rounds 1–8 (B) by experimental conditions and whether the political identity of alter is the same as alter. Margins in B drawn from Model 3 in Table 3. The sample size for Model 3 in Table 3 is 26,053 participant-round-alters. Error bars are 95% confidence intervals computed via bootstrapping.

Fig. 2. Effects of political homophily on severing and proposing ties.

A Marginal effect of political homophily on selecting which alter to drop. Margins drawn from Table S2, Model 2. The sample size for Model 2 in Table S2 is 1251 alters dropped out of 5928 alternatives. Error bars are 95% confidence intervals computed via the Delta method. B Marginal effect of political homophily on which alter to select. Margins drawn from Table S3, Models 1–4. The sample size for Model 1 is 268 selections from 4612 alternatives. The sample size for Model 2 is 380 selections from 7711 alternatives. The sample size for Model 3 is 412 selections from 7711 alternatives. The sample size for Model 4 is 208 selections from 3062 alternatives. Error bars are 95% confidence intervals computed via bootstrapping.

Fig. 3. Changes in networks over time.

Network clustering (A) and segregation (B) at initiation and following each network update by experimental condition. Networks in all conditions became more clustered over time (A). But whether networks became more segregated varied by condition (B).

Fig. 4. How political segregation differs by condition.

Marginal segregation between Democrats and Republicans at the conclusion of the study by experimental condition (A). Margins drawn from Table 4 (N = 40 networks). Error bars are 95% confidence intervals computed via the Delta method. To illustrate, B shows networks in each condition with observed network segregation close to the estimated marginal mean from (A). The networks in B were selected for having segregation patterns typical of the estimated means from (A). They are not representative of other patterns within conditions. For instance, while the Inter/Intragroup network in B has fewer Republicans than the Control and Undifferentiated networks, this is not true of this condition more generally. Indeed, Fig. S25 shows that Republicans are less likely to become isolated from the networks in the Inter/Intragroup condition than in the two other conditions. See Supplementary Note 2 and Figs. S28–S31 of the SI for visualizations of all 40 networks in our experiment, broken down by condition.