Rabin (1993): Intention-Based Reciprocity • The model incorporates three stylized facts: (A) People are willing to sacrifice their own material well-being to help those who are being kind (positive reciprocity). (B) People are willing to sacrifice their own material well-being to punish those who are being unkind (negative reciprocity). (C) Both motivations (A) and (B) have a greater effect on behavior as the material cost of sacrificing becomes smaller (fairness motives are “a normal good”). • Note that this differs from the inequity aversion model, where utility depends only on payoff differences and not on other individuals' actions or intentions. • Evidence from, e.g., the “mini ultimatum game” suggests that payoffs and intentions matter. 235 Battle-of-the-Sexes Game • Consider the following version of a “battle-of-the-sexes game,” with ππ > 0: Player 2 Player 1 Opera Boxing Opera 2X, X 0, 0 Boxing 0, 0 X, 2X • The battle-of-the-sexes game captures a situation in which two players prefer to play either (opera, opera) or (boxing, boxing) rather than not coordinating. Still, player 1 prefers (opera, opera), and player 2 prefers (boxing, boxing). • The players must decide simultaneously without being able to communicate. 236 Nash equilibria in pure strategies Player 2 Player 1 Opera Boxing Opera 2X, X 0, 0 Boxing 0, 0 X, 2X • Assume first that the players’ utility depends only on their actions (standard game theory) and that they care only about their own material payoffs (narrow self-interest) ο two Nash-equilibria in pure strategies: (opera, opera) and (boxing, boxing) • Assume now that utility can depend on players’ actions and directly on beliefs (psychological game theory) and that players have social preferences ο additional “fairness equilibria” can exist 237 Example of a Fairness Equilibrium (1) • Assume that player 1 (“he”) cares not only about his own payoff but, depending on player 2’s (“she”) motives, also about player 2's payoff. ο If player 2 intentionally helps player 1, then player 1 wants to help player 2 ο if player 2 intentionally hurts player 1, then player 1 wants to hurt player 2 • Suppose player 1 believes that 1) player 2 is playing boxing, and 2) player 2 believes that player 1 is playing boxing (ο¨ second-order belief) • Then player 1 concludes that player 2 is choosing an action that helps both players (player 2 playing opera would hurt both players, i.e., also player 2). • Player 1 thus perceives player 2 as neither generous nor mean. 238 Example of a Fairness Equilibrium (2) • As player 2 is not being either generous (kind) or mean (unkind) to player 1, neither stylized fact (A) nor (B) applies. • Thus, player 1 will be neutral about his effect on player 2 and will pursue his material self-interest, which means playing boxing (given his beliefs). • The same argument can be repeated for player 2, so that (boxing, boxing) is an equilibrium: if it is common knowledge that (boxing, boxing) will be the outcome, then each player is maximizing their utility by playing “boxing.” ο (boxing, boxing) is a fairness equilibrium of the game • Recall that (boxing, boxing) is also a conventional Nash equilibrium of the game. • Due to the symmetry of the game, the same reasoning applies regarding the strategy combination (opera, opera), which is both a “fairness equilibrium” and a Nash equilibrium. 239 Example of a Fairness Equilibrium (3) • Now, to see the importance of fairness, suppose player 1 believes that 1) player 2 is playing boxing, and 2) player 2 believes that player 1 is playing opera (second-order belief) • Now, player 1 concludes that player 2 is lowering her own payoff to hurt him and, therefore, feels hostility toward player 2 and wants to harm her. • Player 1 might then be willing to sacrifice his material payoff and play opera rather than boxing. • If both players have a strong enough emotional reaction to each other's behavior, then (opera, boxing) is a fairness equilibrium. (We will derive this formally.) • Note that (opera, boxing) is not a Nash equilibrium of the game. • The players’ expectations are crucial in the fairness equilibrium: player 1’s utility depends not only on his actions but also directly on his beliefs about player 2’s intentions. 240 Model of Intention-Based Reciprocity • Consider a two-player, normal form game with (mixed) strategy sets ππ1 and ππ2 for players 1 and 2, derived from finite pure strategy sets π΄π΄1 and π΄π΄2 . • Let ππππ : ππ1 × ππ2 → β be player ππ’s material payoffs. • A “psychological game” is constructed from the above “material game” by assuming that each player’s subjective expected utility when she chooses her strategy will depend on three factors: 1. her strategy 2. her beliefs about the other player's strategy choice (belief) 3. her beliefs about the other player's beliefs about her strategy (2nd order belief) 241 Notation • Strategies chosen by the two players: ο strategy of player 1: ππ1 ∈ ππ1 ο strategy of player 2: ππ2 ∈ ππ2 • (1st order) beliefs: ο player 2’ beliefs about what strategy player 1 is choosing: ππ1 ∈ ππ1 ο player 1’ beliefs about what strategy player 2 is choosing: ππ2 ∈ ππ2 • 2nd order beliefs: ο player 1’s beliefs about what player 2 believes player 1’s strategy is: ππ1 ∈ ππ1 ο player 2’s beliefs about what player 1 believes player 2’s strategy is: ππ2 ∈ ππ2 242 Kindness function (1) • The “kindness function” ππππ (ππππ , ππππ ) measures how kind player ππ is being to player ππ. • If player ππ believes that player ππ is choosing strategy ππππ , how kind is player ππ being by choosing ππππ ? • By choosing ππππ , player ππ is choosing the payoff pair ππππ ππππ , ππππ , ππππ ππππ , ππππ from the set of all payoffs feasible, given player ππ is choosing strategy ππππ (which player 1 believes she does). • Set of all feasible payoffs: Π ππππ ≡ ππππ ππ, ππππ , ππππ ππππ , ππ • Let ππππβ ππππ be player ππ’s highest payoff in Π ππππ ππ ∈ ππππ • Let ππππβ ππππ be player ππ’s lowest payoff in Π ππππ among the points that are Paretoefficient in π±π± ππππ 243 Kindness function (2) • Let ππππππ ππππ = ππππβ ππππ + ππππβ ππππ 2 be the “equitable payoff” ο Provides a “reference point” against which it is measured how generous player ππ is being to player ππ. • Let ππππππππππ ππππ be the worst possible payoff for player ππ in Π ππππ • Definition 1: Player ππ’s kindness (by choosing ππππ ) to player ππ is given by: ππππ ππππ , ππππ ≡ ππππ ππππ , ππππ − ππππππ ππππ ππππβ ππππ − ππππππππππ ππππ If ππππβ ππππ − ππππππππππ ππππ = 0, then ππππ ππππ , ππππ = 0. 244 Kindness function (3) • In the following, we disregard the case ππππβ ππππ = ππππππππππ ππππ , where player ππ’s payoff is independent of player ππ’s action. • We have ππππ ππππ , ππππ = 0 if and only if player ππ, by choosing ππππ , gives player ππ her equitable payoff. • If player ππ gives player ππ less than her equitable payoff, we have ππππ ππππ , ππππ < 0 ο This can arise either if player ππ takes more than his equitable share on the Pareto frontier or is choosing an inefficient point in Π ππππ . In the latter case, this might be costly for player 1 himself. • If player ππ gives player ππ more than her equitable payoff, we have ππππ ππππ , ππππ > 0 245 Kindness function (4) • Let the function πποΏ½ππ ππππ , ππππ represent player ππ’s beliefs about how kindly player ππ is treating him. • Definition 2: Player ππ’s belief about how kind player ππ is being to him is given by: πποΏ½ππ ππππ , ππππ ≡ ππππ ππππ , ππππ − ππππππ ππππ ππππβ ππππ − ππππππππππ ππππ If ππππβ ππππ − ππππππππππ ππππ = 0, then πποΏ½ππ ππππ , ππππ = 0. • Due to the normalization of the kindness functions, ππππ ππππ , ππππ and πποΏ½ππ ππππ , ππππ are bounded and lie in the interval 1 2 1 −1, 2 . To see this: ππππ = if (i) ππππβ = ππππππππππ , in which case ππππππ takes on its lowest possible value, and (ii) ππππ = ππππβ , which is the kindest move player ππ can make. Likewise, ππππ = −1 if (i) ππππβ = ππππβ , in which case ππππππ takes on its highest 246 possible value, and (ii) ππππ = ππππππππππ , which is the unkindest move player ππ can make. Players’ Utility Function • Each player ππ chooses ππππ to maximize their expected utility ππππ ππππ , ππππ , ππππ = ππππ ππππ , ππππ + ππΜππ ππππ , ππππ ⋅ 1 + ππππ (ππππ , ππππ ) which incorporates a player’s material payoff and their notion of fairness. • These preferences capture reciprocity: ο If player ππ believes that player ππ is unkind, ππΜππ < 0, then player ππ’s wants to choose an action ππππ such that ππππ is low or even negative , i.e., player ππ wants to be unkind to player ππ. ο Likewise, if ππΜππ > 0, then player ππ’s wants to choose an action ππππ such that ππππ is positive and large, i.e., player ππ wants to be kind to player ππ. 247 Fairness vs. Material Well-Being • Importantly, a player will trade off material well-being against reciprocity: if a reciprocal response is too costly in terms of material well-being, player ππ might, for instance, treat a kind player ππ in an unkind way if the material gain from being unkind is sufficiently large. 1 • Recall that the kindness functions are bounded in between −1, . This 2 implies that the relative importance of fairness is diminished if the material payoffs, which are not bounded, become larger. • This captures stylized fact (C): Fairness concerns play a bigger role if material stakes are smaller. • The model does not pin down the exact relative importance of fairness vs. material well-being; it just captures the qualitative (i.e., not quantitative) trade-off. 248 Psychological Nash Equilibrium • In standard game theory, a player’s utility depends only on the actions chosen by all players (which determine the material payoffs). In contrast, in “psychological games,” a player’s utility may also directly depend on the beliefs of players. • A psychological Nash equilibrium (here: fairness equilibrium) is the analog of a standard Nash equilibrium for “psychological games” and imposes the additional condition that all higher-order beliefs (and not only first-order beliefs) must match actual behavior. • Definition 3: The pair of strategies ππ1 , ππ2 ∈ (ππ1 , ππ2 ) is a fairness equilibrium if, for ππ = 1,2, ππ ≠ ππ, (1) ππππ ∈ arg max ππππ (ππ, ππππ , ππππ ) ππ∈ππππ (2) ππππ = ππππ = ππππ . 249 Example of a Fairness Equilibrium (cont., 4) • Coming back to our earlier example, is the hostile outcome (opera, boxing) a fairness equilibrium? • Consistency of beliefs and actions requires: ο ππ1 = ππ1 = ππ1 = opera: player 1’s second-order belief ππ1 (indicating what player 1 believes about player 2’s belief about player 1’s strategy) equals player 2’s belief ππ1 (indicating what player 2 believes about player 1’s strategy) and it equals player 1’s strategy ππ1 . ο ππ2 = ππ2 = ππ2 = boxing: player 2’s second-order belief ππ2 (indicating what player 2 believes about player 1’s belief about player 2’s strategy) equals player 1’s belief ππ2 (indicating what player 1 believes about player 2’s strategy) and it equals player 2’s strategy ππ2 . 250 Example of a Fairness Equilibrium (5) • Would player 1 have an incentive to deviate from playing “opera”? • Player 1’s utility from playing opera is ππ = 0. • To see this, consider first player 1’s material payoff in this equilibrium, which is given by ππ1 ππ1 = opera, ππ2 = boxing = 0. • Consider next, player 1’s belief about how kind player 2 is being to him: • To calculate πποΏ½2 ππ2 = boxing, ππ1 = opera = ππ1 ππ1 ,ππ2 −ππ1ππ ππ1 ππ1β ππ1 −ππ1ππππππ ππ1 , we must derive its components: ο Payoff in equilibrium: ππ1 ππ1 = opera, ππ2 = boxing = 0 ο Highest possible payoff given ππ1 : ππ1β ππ1 = opera = 2ππ ο Lowest possible payoff given ππ1 : ππ1ππππππ ππ1 = opera = 0 251 Example of a Fairness Equilibrium (6) ο Lowest possible payoff among the Pareto-efficient ones given ππ1 : ππ1β ππ1 = opera = 2ππ ο Equitable payoff (reference point for fairness evalution): ππ1β ππ1 =opera + ππ1β ππ1 =opera 2ππ +2ππ ππ = = 2ππ ππ1 ππ1 = opera = 2 2 • Player 1’s belief about how kind player 2 is being to him is thus given by: πποΏ½2 ππ2 = boxing, ππ1 = opera = ππ1 ππ1 , ππ2 − ππ1ππ ππ1 ππ1β ππ1 − ππ1ππππππ ππ1 0 − 2ππ = = −1 2ππ − 0 • Next, we must consider how kind player 1 is to player 2 by choosing ππ1 = opera. 252 Example of a Fairness Equilibrium (7) • Player 1’s kindness (by choosing ππ1 ) to player 2 is given by: ππ1 ππ1 , ππ2 ≡ ππ2 ππ2 , ππ1 − ππ2ππ ππ2 ππ2β ππ2 − ππ2ππππππ ππ2 • To calculateππ1 ππ1 , ππ2 , we must derive its components: ο Payoff in equilibrium: ππ2 ππ2 = boxing, ππ1 = opera = 0 ο Highest possible payoff given ππ2 : ππ2β ππ2 = boxing = 2ππ ο Lowest possible payoff given ππ2 : ππ2ππππππ ππ2 = boxing = 0 ο Lowest possible payoff among the Pareto-efficient ones given ππ1 : ππ2β ππ2 = boxing = 2ππ ο Equitable payoff: ππ1ππ ππ2 = boxing = 2ππ +2ππ 2 = 2ππ 253 Example of a Fairness Equilibrium (8) • Player 1’s kindness to player 2 is thus given by: ππ1 ππ1 , ππ2 ≡ ππ2 ππ2 , ππ1 − ππ2ππ ππ2 ππ2β ππ2 − ππ2ππππππ ππ2 0 − 2ππ = −1 = 2ππ − 0 • We can now derive player 1’s utility in equilibrium: ππ1 ππ1 = opera, ππ2 = boxing, ππ1 = opera = ππ1 ππ1 = opera, ππ2 = boxing + ππΜ2 ππ2 = boxing, ππ1 = opera ⋅ 1 + ππ1 ππ1 = opera, ππ2 = boxing = 0 + −1 ⋅ 1 + −1 = 0 • What would player 1’s utility be by playing boxing? If his utility is not higher by deviating to boxing, he does not have an incentive to do so. 254 Example of a Fairness Equilibrium (9) • By playing boxing, player 1 would increase his monetary payoff and obtain ππ1 ππ1 = boxing, ππ2 = boxing = ππ • How kind would player 1 be to player 2 in this case? ππ2 ππ2 , ππ1 − ππ2ππ ππ2 2ππ − 2ππ =0 ππ1 ππ1 = boxing, ππ2 = boxing ≡ β = ππππππ 2ππ − 0 ππ2 ππ2 − ππ2 ππ2 • Player 1 would thus be neither kind nor unkind. • Player 1’s utility when deviating (from playing opera to playing boxing): ππ1 ππ1 = boxing, ππ2 = boxing, ππ1 = opera = ππ1 ππ1 = boxing, ππ2 = boxing + ππΜ2 ππ2 = boxing, ππ1 = opera ⋅ 1 + ππ1 ππ1 = boxing, ππ2 = boxing = ππ + −1 ⋅ 1 + 0 = ππ − 1 • Note that we consider a possible deviation from equilibrium, thereby keeping constant player 2’s actions and beliefs, and thus fairness. 255 Example of a Fairness Equilibrium (10) • That is, if ππ < 1, i.e., if the monetary payoffs are relatively small, player 1 does not want to deviate from playing opera (given his beliefs and second-order beliefs) ο¨ he prefers to forgo money to reciprocate player 2’s unkind intentions. • Since the game is symmetric, the same reasoning applies to player 2. • This shows that (opera, boxing) is a fairness equilibrium of the game for sufficiently small monetary payoffs (ππ < 1). • By symmetry: (boxing, opera) is also a fairness equilibrium if ππ < 1. • Since the kindness functions take on value 0 in case the two players coordinate, both “classic” Nash equilibria in pure strategies, (opera, opera) and (boxing, boxing), are also fairness equilibria of the game. • The interesting implication of the model are the additional equilibria, showing that fairness motives can explain why the players are willing to impose costly punishment on each other in equilibrium. 256 Fairness Equilibria in the Prisoner’s Dilemma (1) • The analysis of the battle-of-the-sexes game exemplified “the dark side of fairness,” which can lead to hostility in equilibrium. • By contrast, fairness may also lead each player to sacrifice material payoffs to help the other player, which we will see in the analysis of the prisoner’s dilemma. Player 2 Player 1 Cooperate Defect Cooperate 4X, 4X 0, 6X Defect 6X, 0 X, X 257 Fairness Equilibria in the Prisoner’s Dilemma (2) • There is a unique Nash equilibrium in dominant strategies in this game: (defect, defect). • (A dominant strategy is a strategy that is optimal for a player irrespective of the strategies of other players.) • Can the outcome (cooperate, cooperate) be sustained as a fairness equilibrium? • Suppose (cooperate, cooperate) is an equilibrium. • Consider first player 1. Would he have an incentive to deviate from playing cooperate? • Player 1’s utility from playing cooperate is ππ = 4ππ + 3 . 4 • To see this, consider first player 1’s material payoff in equilibrium, which is given by ππ1 ππ1 = cooperate, ππ2 = cooperate = 4ππ. 258 Fairness Equilibria in the Prisoner’s Dilemma (3) • Consider next, player 1’s belief about how kind player 2 is being to him: • To calculate πποΏ½2 ππ2 = coop, ππ1 = coop = ππ1 ππ1 ,ππ2 −ππ1ππ ππ1 ππ1β ππ1 −ππ1ππππππ ππ1 , we must derive its components: ο Payoff in equilibrium: ππ1 ππ1 = coop, ππ2 = coop = 4ππ ο Highest possible payoff given ππ1 : ππ1β ππ1 = coop = 4ππ ο Lowest possible payoff given ππ1 : ππ1ππππππ ππ1 = coop = 0 ο Lowest possible payoff among the Pareto-efficient ones given ππ1 : ππ1β ππ1 = coop = 0 (note that the payoff pair (0,6X) is Pareto-efficient) ο Equitable payoff: ππ1β ππ1 =coop + ππ1β ππ1 =coop 4ππ +0 ππ ππ1 ππ1 = coop = = = 2ππ 2 2 259 Fairness Equilibria in the Prisoner’s Dilemma (4) • Player 1’s belief about who kind player 2 is being to him is thus given by: πποΏ½2 ππ2 = coop, ππ1 = coop = ππ1 ππ1 , ππ2 − ππ1ππ ππ1 4ππ − 2ππ 1 = = 4ππ − 0 2 ππ1β ππ1 − ππ1ππππππ ππ1 • Next, we consider how kind player 1 is being to player 2 by choosing ππ1 = coop. • To calculateππ1 ππ1 , ππ2 , we must derive its components: ο Payoff in equilibrium: ππ2 ππ2 = coop, ππ1 = coop = 4ππ ο Highest possible payoff given ππ2 : ππ2β ππ2 = coop = 4ππ ο Lowest possible payoff given ππ2 : ππ2ππππππ ππ2 = coop = 0 ο Lowest possible payoff among the Pareto-efficient ones given ππ1 : ππ2β ππ2 = coop = 0 ο Equitable payoff: ππ1ππ ππ2 = coop = 4ππ +2ππ 2 = 2ππ 260 Example of a Fairness Equilibrium (5) • Player 1’s kindness to player 2 is thus given by: ππ1 ππ1 , ππ2 ≡ ππ2 ππ2 , ππ1 − ππ2ππ ππ2 ππ2β ππ2 − ππ2ππππππ ππ2 4ππ − 2ππ 1 = = 4ππ − 0 2 • We can now derive player 1’s utility in equilibrium: ππ1 ππ1 , ππ2 , ππ1 = ππ1 ππ1 , ππ2 + ππΜ2 ππ2 , ππ1 ⋅ 1 + ππ1 ππ1 , ππ2 1 1 3 4ππ + ⋅ 1 + = 4ππ + 2 2 4 = • Would player 1 have an incentive to deviate from playing cooperation? 261 Fairness Equilibria in the Prisoner’s Dilemma (6) • By playing defect, player 1 would increase his monetary payoff and obtain ππ1 ππ1 = defect, ππ2 = coop = 6ππ • How kind would player 1 be to player 2 in this case? ππ2 ππ2 , ππ1 − ππ2ππ ππ2 0 − 2ππ 1 =− ππ1 ππ1 = defect, ππ2 = coop ≡ β = ππππππ 4ππ − 0 2 ππ2 ππ2 − ππ2 ππ2 • Player 1 would thus be unkind. • Player 1’s utility when deviating (from playing cooperation to playing defect): ππ1 ππ1 = defect, ππ2 = coop, ππ1 = coop = ππ1 ππ1 = defect, ππ2 = coop + ππΜ2 ππ2 = coop, ππ1 = coop ⋅ 1 + ππ1 ππ1 = defect, ππ2 = coop = 1 1 1 6ππ + ⋅ 1 − = 6ππ + 2 2 4 262 Fairness Equilibria in the Prisoner’s Dilemma (7) • Deviation from cooperate to defect yields a lower utility for player 1 if 1 3 6ππ + < 4ππ + 4 4 1 ππ < 4 • Since the game is symmetric, the same reasoning applies to player 2. • The strategy pair (cooperate, cooperate) is thus a fairness equilibrium if the monetary payoffs are sufficiently small (ππ < 1/4). • In the fairness equilibrium (cooperate, cooperate), each player wants to help the other player by playing cooperate, thereby forgoing material gains from defecting. • It can be shown that the conventional Nash equilibrium (defect, defect) is also a fairness equilibrium of the game. 263 Reciprocity vs. Pure Altruism • The concept of pure, or unconditional, altruism cannot explain that both (defect, defect) and (cooperate, cooperate) can be outcomes of the game. • To see this, consider the following: 1) If player 1 thought that player 2 was playing cooperate, he would be paying cooperate if he was willing to give up 2ππ to increase player 2’s payoff by 4ππ. 2) If player 1 thought that player 2 was playing defect, he would be paying cooperate if he was willing to give up ππ to increase player 2’s payoff by 5ππ. • Hence, if player 1 cooperates in response to cooperation (point 1), he would cooperate in response to defect (point 2). • The concept of pure altruism thus cannot explain both equilibria, though both outcomes seem plausible. 264 Importance of Non-Chosen Strategies (1) • In the Mini Ultimatum Game, we have seen the importance of non-chosen strategies (i.e., the respective alternatives to allocation 8:2). • To see this point in the context of the Rabin-model, consider the following game, labelled “Prisoner’s Non-Dilemma.” Player 2 Cooperate Player 1 Cooperate 4X, 4X Defect 6X, 0 • Player 2 has no choice but must play cooperate. 265 Importance of Non-Chosen Strategies (2) • In the Prisoner’s Non-Dilemma, (cooperate, cooperate) is no fairness equilibrium. • To see this, consider that player 2 cannot be kind (she cannot be unkind either) because she cannot take an action. • Since πποΏ½2 = 0, player 1 will maximize his material well-being and choose to defect. • It follows that the unique fairness equilibrium (and unique Nash equilibrium) is given by (defect, cooperate). • The reason behind (cooperate, cooperate) being no fairness equilibrium in the Prisoner’s Non-Dilemma is that player 2 does not have “defect” in her strategy space and thus cannot intentionally forgo choosing defect, thereby being kind to player 1. 266 Relation of Fairness and Nash Equilibria (1) • In the examples discussed so far, all Nash-equilibria were also fairness equilibria, but not all fairness equilibria were Nash-equilibria. • This is not a general property of the model, as the following “chicken game” exemplifies: Player 2 Dare Chicken Player 1 Dare Chicken -2X, -2X 2X, 0 0, 2X X, X • The two Nash-equilibria in pure strategies are (dare, chicken) and (chicken, dare). • Are these also fairness equilibria of the game? 267 Relation of Fairness and Nash Equilibria (2) • Consider the Nash-equilibrium (dare, chicken). • Given this strategy combination, player 1 is unkind to player 2 because he could choose “chicken” and thereby increase player 2’s payoff from 0 to ππ. • Player 2 would thus want to be unkind herself (negative reciprocity) and choose “dare,” provided the cost in terms of material well-being are not too large. • Hence, for sufficiently small ππ, (dare, chicken) is not a fairness equilibrium of the game, even though it is a Nash-equilibrium. • Due to the symmetry of the game, the same applies to (chicken, dare). • This exemplifies that Nash-equilibria can exist that are not consistent with fairness motives. • The paper by Rabin (1993) derives general properties of the model, and discusses some applications and limitations, all of which is left as optional reading. 268