Presentation on theme: "In Broad Daylight: Fuller Information and Higher-Order Punishment Opportunities Can Promote Cooperation Kenju Kamei – Bowling Green State University Louis."— Presentation transcript:
In Broad Daylight: Fuller Information and Higher-Order Punishment Opportunities Can Promote Cooperation Kenju Kamei – Bowling Green State University Louis Putterman – Brown University
In laboratory experiments with many subject pools, voluntary contributions to a group activity (public good) can be stabilized by opportunities to assign costly punishment (informal sanctions [IS]). But some second-generation contribution- and-punishment experiments have suggested that an overly sanguine impression about IS arises from artificial suppression of counter- punishment and vendettas.
We investigate experimentally whether providing subjects with full information on initial punishments and allowing unrestricted higher-order punishment reduces the benefits of IS in a VCM setting. We reconfirm that counter-punishment is efficiency-reducing in an information-poor environment like that of Nikiforakis (2008), but we find that generalized higher-order punishment opportunities in a fuller- information environment are associated with higher, not lower, cooperation and earnings.
Background A large literature (Ledyard, 1995; Zelmer, 2003; Gӓchter and Herrmann, 2008) shows that in the finitely repeated VCM with n 2 and 0 < mpcr < 1, contributions decline with repetition (absent punishment, communication, etc.). Fehr and Gӓchter (2000) and numerous subsequent studies showed that adding a second stage to each period, in which subjects can engage in costly IS, reverses or retards the decline.
A figure from Fehr and Gӓchter (2000).
To prevent vendettas, FG randomly scrambled subject IDs in their partner groups; even so, there was some punishment of high contributors, which Cinyabuguma, Page and Putterman (2006) and Herrmann, Thöni and Gӓchter (2008) argue may mainly be attempts to counter-punish. (Herrmann et al. find that in subject pools of countries in southeast and eastern Europe and the middle east, anti- social punishment is so common that availability of IS doesnt aid cooperation.)
Nikiforakis (2008) suggested that identification of ones punisher and the possibility to retaliate is more realistic in many situations. He replicated FG (2000)s results in one treatment but in another added a third stage in which subjects could counter-punish. In Nikiforakiss counter-punishment treatment, a subject i learns only who punished i by how much and can only counter-punish those individuals (in a dedicated 3 rd stage). After the 3 rd stage, a new period begins with scrambled identities.
Nikiforakis (2008) partner treatment results.
Similar results obtained by Hopfensitz and Reuben (2009) and Engel, Kube & Kurschilgen (2011). Denant-Boemont, Masclet and Noussair (2007) replicated Nikiforakiss counter-punishment results, also introduced a fuller-information treatment. In it, each period has 3 stages as in Nikiforakis (2008) but subjects are informed of all stage 2 bi- lateral punishments and anyone can engage in stage 3 punishment. In principle, both counter-punishment and punishment enforcement are possible.
Punishment enforcement can include - Punishing individuals who fail to punish low contributors to the group account + - Punishing individuals who punish anti- socially* or perversely** ++ DB-et al.s fuller information treatment produced cooperation levels less harmed by availability of the 3 rd stage (2 nd punishment opportunity) than the counter-punishment only treatment, but no improvement in cooperation or efficiency relative to FG (2000). + PEO punishment enforcement for omission ++PEC punishment enforcement for commission
Punishment enforcement can include - Punishing individuals who fail to punish low contributors to the group account + - Punishing individuals who punish anti- socially* or perversely** ++ DB-et al.s fuller information treatment produced cooperation levels less harmed by availability of the 3 rd stage (2 nd punishment opportunity) than the counter-punishment only treatment, but no improvement in cooperation or efficiency relative to FG (2000). + PEO punishment enforcement for omission ++PEC punishment enforcement for commission * Anti-social punishment: i punishes j, where C i < C j. ** Perverse punishment: a contributor of more than the group average for the period is punished. Most cases of anti-social punishment are also cases of perverse punishment, but this isnt necessary. Consider: contributions 2, 4, 6, 8. Contributor of 2 punishes contributor of 4. This punishment is anti- social but not perverse.
Punishment enforcement can include - Punishing individuals who fail to punish low contributors to the group account + - Punishing individuals who punish anti- socially* or perversely** ++ DB-et al.s fuller information treatment produced cooperation levels less harmed by availability of the 3 rd stage (2 nd punishment opportunity) than the counter-punishment only treatment, but no improvement in cooperation or efficiency relative to FG (2000). + PEO punishment enforcement for omission ++PEC punishment enforcement for commission * Perverse punishment: a contributor of more than (or equal to) the group average for the period is punished. ** Anti-social punishment: i punishes j, where C i < C j. Most cases of anti-social punishment are also cases of perverse punishment, but this isnt necessary. Ill call punishment that is not perverse: normal.
Cinyabuguma et al. (2006) studied higher-order punishment (punishment conditioned on others 1 st order punishing) in a more constrained setting and found efficiency to be just as high with such punishment avialable as in a standard FG type cooperation-and-punishment experiment. Recent experiments on feuding allow the number of punishment stages in a period to be determined endogenously. These include Nikiforakis and Engelmann (2011), Nicklisch and Wolff (2011), and Nikiforakis, Noussair and Wilkening (forthcoming).
Motivation of Our Experiment We conjecture that fuller information about who punished whom may improve cooperation both (a) by hastening norm emergence, and (b) by making possible punishment enforcement (especially Punishment Enforcement for Commission [PEC] rather than Punishment Enforcement for Omission [PEO]).
Motivation (cont.) We suspect that dedicated higher-order punishment stages can have an experimenter demand effectif subjects have no other decisions to make except to punish again, they are more likely to punish again. We therefore limit the number of stages available for nothing other than punishing, instead letting punishment of high order occur by maintaining identity information across periods. Punishment opportunities and the (1 st order) social dilemma itself (how much to contribute) are thus inter-leaved in terms of action stages.
Experimental Design min
Design (cont.) Subjects play for 15 periods in partner groups. There are two Reference treatments without possibility of (informed) higher-order punishment, six treatments with h-o pun. 3 EGO treatments: Subject IDs change each period, subjects only learn who punished them, subjects can only counter-punish. 3 FULL treatments: Subject IDs fixed for all 15 periods, subjects learn all bilateral punishments, subjects can engage in higher-order punishment of any group member(s).
Design (cont.) In each category (EGO, FULL) there is a 2-stage treatment, a 3-stage treatment without additional history information, and a 3-stage treatment with additional history information: EGO2, EGO3, EGO3hist FULL2, FULL3, FULL3hist The purpose of 2-stage treatments is to reduce experimenter demand, but at least last periods punishment must be shown at this periods punishment stage, to permit counter-punishment. We have 3hist treatments to have both history display (as in 2-stage treatments) and the extra punishment stage (as in EGO3 and FULL3).
Design (cont.) For comparison, each set of treatments has its Reference treatment: EGO2, EGO3, EGO3hist, Reference (random ID) FULL2, FULL3, FULL3hist, Reference (fixed ID) Reference treatments have 2 stages per period and no possibility of informed higher-order punishment (similar to FG2000, etc.).
Contributions are significantly higher in FULL3 and in FULL3hist than in Ref (fixed) Contributions are significantly lower in EGO3 than in Ref (random) and EGO2.
Earnings are significantly higher in FULL3hist than in Ref (fixed). Earnings are significantly lower in EGO3 than in Ref (random), higher in EGO3hist than in EGO3, and higher in EGO2 than in EGO3.
The one treatment that performs significantly worse than others is EGO3, which closely parallels the Nikiforakis (2008) counter-punishment treatment. Either eliminating a 3 rd stage dedicated to punishment or carrying over history from period to period suffice to reduce inefficiency relative to the Reference treatments. (Both not needed.) The treatment that performs best is FULL3hist, with fullest information and higher-order punishment opportunities.
There is more cost from punishment in the Reference treatments than in corresponding treatments with higher-order punishment opportunities, except for the EGO3 vs. Reference (random) comparison. Expenditure on, and loss of earnings to, punishment
Feuding experiments such as Nikiforakis and Engelmann (where presence of higher order punishment opportunities is less damaging to contributions and earnings than in Nikiforakis  but definitely not helpful) provide equally full information and opportunities to engage in higher-order punishment of anyone, as in our FULL treatments. The difference is that in N&E11 punishing continues in each period until a stage is reached at which no one wants to punish or no one can punish or can be punished (due to budget depletion).
In our treatments, theres a maximum of two punishment stages before the next period brings a new contribution stage. Which approach (limited vs. endogenous punishment stages) is more realistic may depend on what environment is being modeled. We think our approach makes sense for understanding an ongoing collective activity, e.g. a work team.
What explains poor performance in EGO treatments, especially EGO3, vs. good performance in FULL treatments, especially FULL3hist? We find little evidence for punishment enforcement of either type (PEO, PEC) in the FULL treatments. But we find different patterns of counter- punishment in EGO vs. FULL treatments.
Perverse and normal 1 st order punishers are about equally likely to be counter-punished, and receive similar amounts of counter- punishment, in the EGO treatments, especially EGO3, whereas Perverse 1 st order punishers receive significantly more counter-punishment than normal 1 st order punishers in the FULL treatments, suggesting more agreement there on what (1 st order) punishments are justified.
Regressions to predict 3 rd stage (2 nd order) punishment as a function of contributions and 2 nd stage (1 st order) punishments in the same period show: - significant correlations between 1 st order punishment and 2 nd order counter-punishment amounts in the FULL but not EGO treatments - significantly more counter-punishment of perverse than of normal 1 st order punishers per unit of 1 st order punishment in the FULL treatments
Dep. Var.: 3 rd stage (2 nd order) pun. i gave to j (We control for determinants of 1 st stage pun.) Main expected determinants of 2 nd stage pun.
We see the take-away message being that in the FULL treatments, especially FULL3hist, theres more concerted counter- punishment of perverse punishers, which perhaps helps drive higher contributions and efficiency in those treatments.
In both 2 stage treatments (EGO2, FULL2), regressions suggest significant counter- punishment of 1 st order punishments of the previous period, with greater counter- punishment of perverse than of normal 1 st order punishment in both treatments. (Note that teasing apart punishing for contribution choice from punishing for past punishment choice is not an exact process.)
Counter- punishment of normal punishers Counter- punishment of perverse punishers.
In both 2 stage treatments (EGO2, FULL2), regressions suggest significant counter- punishment of 1 st order punishments of the previous period, with greater counter- punishment of perverse than of normal 1 st order punishment in both treatments. Thus, reducing separation of punishment and counter-punishment opportunities doesnt eliminate counter-punishment, but the observed counter-punishment doesnt interfere with and may actually reinforce the ability of punishment to help sustain cooperation.
We also estimated regressions with treatment dummy variables to see how our four treatment parameters affect contributions and earnings respectively. The parameters are: I – Identities remain fixed across periods T – Third Stage H – History from previous periods shown F – Full information shown on all punishments and opportunities to punish any subject.
Results show -Having fixed identifiers per se makes no significant difference to contributions or earnings -Having a 3 rd stage significantly reduces both contributions and earnings. -However, having full information along with 3 rd stage significantly increases contributions and earnings. On net, contributions are slightly increased by having both T and F, earnings are slightly reduced by having both T & F. -Having history information per se increases earnings by a smaller amount than the effects of 3 rd stage or T & F, and with less statistical significance. These results should be treated as suggestive only because they use individual level observations (averaged for an individual over all 15 period) without clustering errors by group. When errors are clustered or when group level observations are used in the regressions, there are no statistically significant results.
Conclusion Our data suggest that the danger that one will be counter-punished if punishment recipients learn who punished them and how much doesnt necessarily undermine the ability of IS to promote cooperation. In several treatments, the availability of the information and opportunities for higher-order punishment are associated with higher contributions than in treatments with opportunities for 1 st order punishment only. In one case, the difference is statistically significant.
We find little evidence of (3 rd party) punishment enforcement. Instead, treatments with fuller information and more general opportunities for higher-order punishment are marked by more pro-social bias in the pattern of counter-punishment: those who 1 st order punish perversely receive more counter- punishment relative to non-perverse 1 st order punishers in these FULL treatments.
We conjecture that more complete information about the pattern of punishment in the group favors the emergence of norms supportive of cooperation (when the subject pool is one in which such norms are likely to be supported). This result is consistent with the typical presence of more cooperatively-inclined than anti-socially inclined individuals in the subject pool.
Seeing that low contributors are punished more than high ones are and that punishers of high contributors tend to be heavily punished may lead more subjects to cooperate more in the contribution stage and to punish pro- socially if at all. The idea that voluntary cooperation or decentralized collective action can be aided by punishment opportunities may accordingly survive the availability of counter-punishment.
But the danger of feuds appears to discourage initial punishment and even rare feuds deplete resources in the design of Nikiforakis and Engelmann (2011). Conclusions may therefore turn on which experimental setting is more relevant to which real world problems of interest.