Networks and Cooperation John T. Scholz Florida State University Political Networks Workshop May 19, 2009 Background: Why Networks? • Local Institutions develop to coordinate policies where environmental problems are most acute. – Mark Lubell, John Scholz, Mark Schneider, and Mihriye Mete. 2002. “Watershed Partnerships and the Emergence of Collective Action Institutions” American Journal of Political Science 46: 148-163. • Federal policies can enhance the capacities of local policy networks to integrate policies within these institutions • Mark Schneider, Mark Lubell, John T. Scholz, Denisa Midruta, and Matt Edwards. 2003. “Building Consensual Institutions: Networks and the National Estuary Program” American Journal of Political Science 47:143-158. Background: Why Networks? • Local institutions and policy networks transform interests and strongly influence independent federal policies – John T. Scholz and Cheng-Lung Wang. 2006. “Cooptation or Transformation? Local Policy Networks and Federal Regulatory Enforcement” American Journal of Political Science 50(1): 81-97 Bonding Social Capital A’s Network Position Measures Degree 3 Clustering 1 Centrality 0 Bridging Social Capital B’s Network Position Measures Degree 3 Clustering 0 Centrality 0.8 The Risk Hypothesis • Bridging relationships most effective for low risk coordination/assurance games. – Efficient information transmission • Bonding relationships most effective for higher risk dilemmas, like prisoners dilemmas – Norms and 3rd party punishment Two questions for networks • The impact of networks on cooperative behavior (performance) – Network structures and general level of cooperation – Network positions and propensity to cooperate. • The dynamics of network selection in dilemmas – When do actors seek bridging, when bonding? Bridging in Estuary Policy Arenas • Betweenness centrality of ego, not clustering, enhances level of collaboration – John T. Scholz, Ramiro Berardo, and Brad Kile. 2008.“Do Networks Solve Collective Action Problems? Credibility, Search, and Collaboration” Journal of Politics 70(2):393-406 • Actors seek central coordinators, not transitivity – Different organizations are central – Indicates dynamic preference for popular actors – Ramiro Berardo and John T. Scholz, “Self-Organizing Policy Networks: Risk, Partner Selection and Cooperation in Estuaries” 2009? Bonding Capital: 3 Studies • Learning Networks – Evolution of cooperation on fixed networks – Agent-based computational model • Game Networks: Repeated Prisoners Dilemma – Experiment 1: Network impact on performance: • Bridging vs Bonding Capital, fixed networks – Experiment 2: Endogenous networks and performance Learning Networks and Cooperation • Evolution transforms dilemma into coordination game of strategy selection • TFT replaces AllD in finite populations, if – cooperative payoffs and repetitions before updating favor cooperation when pTFT>1/3 – : ρ > 1/N >ρ – ρ depends on TFT AllD • types of strategies (retaliatory, altruistic, exploitative) • clustering of contact types (replaces N) • Learning/updating algorithm and learning network From: Nowak et al. 2004. Emergence of cooperation and evolutionary stability in finite populations Nature 428: 646-50 Learning Networks and Cooperation • “Five Rules for the Evolution of Cooperation” – – – – – Kinship Direct Reciprocity (Repetition) Indirect Reciprocity (Reputation) Network Reciprocity (Structure of Relationships) Group Selection Martin A. Nowak, 2006; Science 314: 1560 Learning Networks and Cooperation • Graph game (overlapping k-person public goods game): – – Cooperators give benefit b to all k neighbors, at cost c Defectors get benefits, do not give, and incur no cost. • Selection favors Cooperation if b/c > k (Ohtsuki et al Nature 2006) Learning Networks and Cooperation • Fixed game structure (10 regulators play all firms) • Vary learning structures • Ring, random, small world, and village • Size: N ={25, 50, 100, 200} • Connectivity: Degree= {2,4,8} and {5,10} • Proportionate Learning algorithm • Copy best response among friends, with innovation • All 1-period strategies • Together, they define a Markov chain • Invariant distribution determines cooperation Learning Networks and Cooperation Learning Networks and Cooperation Learning Networks and Cooperation Learning Networks and Cooperation • Testable Hypotheses: Cooperation increases with: – Bridging Capital (faster exploitation of TFT): • Larger networks (to some threshold) • More connected (higher degree) – Bonding Capital (better defense against AllD) • Higher clustering – Greatest impact in small, less connected networks – Particularly effective in concentrated village structure II. Do Closed Triads Increase Cooperation? • Experimental Design – 66 subjects (in 3 separate sessions) – complete survey first (measure trust and trusting behavior) – assigned randomly in groups of 3 – Play 20 periods of PD (end known in advance) • Payoffs: 0, 25, 75, 100 – Reassign, repeat for 4 complete rounds of 20 periods each Treatment: Closed vs Open Triad Bridging Open Leader= A Open Follower = B, C Bonding Experimental Design Logit analysis of cooperation Variable Coefficient Std. Error Constant:OpenNonleader 0.88 0.19 Closed 0.11 0.07 0.27** 0.10 0.09 0.11 0.07 0.11 Trust Behavior X Closed 0.20** 0.05 T rustEnvironment 0.20** 0.02 Closedfirst 0.00 0.25 round2 0.77* 0.08 round3 0.82* 0.08 round4 0.33* 0.08 Last 3 periods -1.46 0.07 Open Leader Trust: Behavior Attitude Predicted Probability of Cooperation Predicted Probability of Cooperation Do Closed Triads Increase Coordination? • What comparable game can represent the coordination problem? – Symmetric: Pick same alternative from multiple choices • Follow the leader: Open will coordinate in 2 periods – Asymmetric: • Open leader can exploit, but how much? • Closed will do better if compromise (focal) solution exists? III. Do Cooperators Cluster? Does Clustering Increase Cooperation • Market for Lemons: – Low value exchanges replace high value when information asymmetry induces opportunism • Exchange: baubles replace diamonds • Collaboration: low risk replaces high risk, high gains – Represented by PD payoffs, where • CC= high value exchange (diamonds for diamonds) • DD= low value exchange (glass beads for beads) • CD, DC = opportunism (beads for diamonds) Market for Lemons • Voluntary exchanges: – Select partner only when both agree – Execute the exchange – Drop partner whenever dissatisfied Lemons Experiment • 14 subjects per session, computer assisted. • Subjects propose, • When proposals match, play pd with each partner. (Increasing costs for multiple matches) • Repeat for 20 rounds (known in advance). • Reputation mechanism manipulation: – Experiential – Local – global Benchmarks and Conjectures • Pessimists: Stage game payoff dominant Nash equilibrium – 4 links per person, Defect in all games – Earning: 60.4 ECU’s • Optimists: Social Optimum – 10 links per person, Cooperate in all games – Earning: 393.6 ECU’s. • Behavioral Conjectures – Selection: Optimists will cluster – Influence: Clustered player will learn to cooperate 0.4 50 0.2 100 150 200 250 300 Average Earnings Baseline Central Local 0.6 0.8 Average Earnings 0 0.0 Average % of C Plays 1.0 Average % of C Plays 5 10 15 20 5 10 15 Average # CC Links Average # of Links 6 4 2 Average # of Links 3 2 0 1 0 Average # CC Links 4 8 Period 5 Period 20 5 10 Period 15 20 5 10 Period 15 20 Optimists Banish Pessimists to Nashville • A movie is worth 10000 words The Optimists’ Advantage • Optimists play Quit-for-Tat (QFT) – Voluntary dilemma extension of TFT – Cooperate with all new links – Don’t associate with any defectors (cut link) • In good context (many QFT) optimists cluster together to trade diamonds, not beads – all earn socially optimal payoff • As context deteriorates, QFT becomes more generous – Maintain links, play TFT before cutting tie The Pessimists’ Disadvantage • Pessimists play Defect, don’t quit (DnoQ) – Cluster of DnoQ maintains trade in beads, not diamonds – Cut off from QnoT at first encounter – Earn “nash” payoff with lower earning (25 vs 75) – Supports fewer links (optimal at 4 links) – Expected payoff per period in experiment • QFT vs QFT= ???? 100 • DnoQ vs DnoQ= 400 QFT versus DnoQ Strategy Number observed Total Earning QFT 8 5967 Other 42 3970 DnoQ 6 3276 All 56 4181 Strategies categorized by actions in first 3 periods Comparison of cooperation decisions between top and bottom earners (local) Cooperate After Periods 1-3 Periods 15-17 Bottom 25% Top 25% Bottom 25% Top 25% No link 0.38 0.92 0.29 0.30 CC 0.69 1.00 0.87 1.00 DD 0.11 0.00 0.08 0.00 CD 0.33 0.71 0.42 0.00 DC 0.33 1.00 0.13 . Comparison of linking decisions between top and bottom earners (local) Link After Periods 1-3 Periods 15-17 Bottom 25% Top 25% Bottom 25% Top 25% No link 0.36 0.46 0.46 0.15 CC 0.93 0.97 1.00 1.00 DD 0.71 0.75 0.79 0.38 CD 0.64 0.39 0.88 0.33 DC 0.67 0.83 0.89 0.00 Conclusion: Self-organizing resolution of the market for lemons • Optimists ban pessimists to Nashville – Self-organizing diamond exchange: cooperators cluster together and earn more. – Bead exchange banished to “Nashville”. • Explanation: optimists play QFT and pessimists play DnoQ. – Subjects playing QFT earn much more than those playing DnoQ – Top earners tend to play QFT