Charles University in Prague Faculty of Social Sciences Institute of Economic Studies BACHELOR THESIS Does Payoff Dominance Matter? An Experiment. Author: Michal Polena Supervisor: PhDr. Lubomı́r Cingl Academic Year: 2013/2014 Declaration of Authorship The author hereby declares that he compiled this thesis independently, using only the listed resources and literature. The author grants to Charles University permission to reproduce and to distribute copies of this thesis document in whole or in part. Prague, May 15, 2014 Signature Acknowledgments I would like to express my gratitude to my excellent supervisor PhDr. Lubomı́r Cingl for his numerous helpful comments and advices. He has immensely contributed to my academic development. Moreover, I would like to thank my mother and my grandparents for their support during the writing of my bachelor thesis as well as during my entire life. Bibliographic Entry Polena, M. (2014): ”Does Payoff Dominance matter? An experiment.” (Unpublished bachelor thesis). Charles University in Prague. Supervisor: PhDr. Lubomı́r Cingl Length 61 958 characters Abstract Risk dominance and Payoff dominance are considered to be the most important selection criteria in Stag-hunt games. In contrast, the main finding of Schmidt et al. (2003) is that players do not respond to changes in Payoff dominance parameter in these games. There might be, however, other explanations for results of Schmidt et al. (2003). Moreover, Dubois et al. (2011) and Battalio et al. (2001)’s experimental results suggest that sufficiently large changes in Payoff dominance parameter may play a role. We, therefore, proposed three Stag-hunt games in order to examine whether players respond to large changes in Payoff dominance parameter. Furthermore, we tested the predictive power of Relative riskiness. Our main finding is that even large changes in Payoff dominance parameter do not induce players to change their choices. An insignificant trend in players’ choices, caused by Relative riskiness, was detected in our second finding. Possible explanations are discussed. JEL Classification Keywords C72, C92, D81 Game theory, Stag-hunt game, Selection criteria, Payoff dominance, Relative riskiness Author’s e-mail Supervisor’s e-mail [email protected] [email protected] Abstrakt Risk dominance a Payoff dominance jsou pokládána za nejdůležitějšı́ výběrová kritéria v hrách lov na jelena. Naproti tomu, hlavnı́ závěry ze studie Schmidt et al. (2003) jsou, že změny v Payoff dominance parametru neovlivňujı́ hráče v těchto hrách. Nicméně existujı́ dalšı́ možná vysvětlenı́ pro výsledky studie Schmidt et al. (2003). Navı́c, experimentálnı́ výsledky Dubois et al. (2011) a Battalio et al. (2001) naznačujı́, že dostatečně velké změny v Payoff dominance parametru mohou hrát roli. Z tohoto důvodu jsme navrhli tři hry lov na jelena, abychom prozkoumali, jestli hráči reagujı́ na dostatečně velké změny v Payoff dominance parametru. Dále jsme otestovali prediktivnı́ sı́lu Relative riskiness. Našı́m hlavnı́m nálezem je, že ani velké změny v Payoff dominance parametru nepřimějı́ hráče ke změně jejich voleb. Nesignifikantnı́ trend v hráčských volbách způsobený dı́ky Relative riskiness byl objeven v našem druhém nálezu. V práci jsou diskutována možná vysvětlenı́. Klasifikace JEL Klı́čová slova C72, C92, D81 Teorie her, Lov na jelena hra, Výběrová kritéria, Payoff dominance, Relative riskiness E-mail autora [email protected] E-mail vedoucı́ho práce [email protected] Contents List of Tables viii List of Figures ix Thesis Proposal x 1 Introduction 1 2 Game Theory 2.1 A Brief Introduction to Game Theory 2.2 Short Historical Overview . . . . . . 2.3 Game Theory Definitions . . . . . . . 2.4 Classification of Games . . . . . . . . 3 3 4 5 9 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 Stag-hunt Game 12 3.1 An Introduction to Stag-hunt games . . . . . . . . . . . . . . . 12 3.2 Stag-hunt Game in the Perspective of Game Theory . . . . . . . 13 3.3 Structure of Stag-hunt Game . . . . . . . . . . . . . . . . . . . 14 4 Selection Criteria & Parameters 4.1 Payoff Dominance Criterion . . . . 4.2 Risk Dominance Criterion . . . . . 4.3 Optimization Premium . . . . . . . 4.4 Relative Riskiness . . . . . . . . . . 4.5 Mixed Strategies & Mixed Minimax . . . . . . . . . . . . . . . . Regret . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 17 18 19 19 20 5 Literature Survey 21 6 Experiment Design 26 Contents vii 7 Experiment 29 7.1 Experiment Procedure . . . . . . . . . . . . . . . . . . . . . . . 29 7.2 Hypothesis Testing . . . . . . . . . . . . . . . . . . . . . . . . . 30 7.3 Further Data Analyses . . . . . . . . . . . . . . . . . . . . . . . 32 8 Conclusion 35 Bibliography 39 A Instructions I B Experimental screen IV C Output from Stata VI List of Tables 7.1 7.2 7.3 7.4 Demographic Data of Participants Experimental results . . . . . . . Participants’ Types . . . . . . . . Coordination & Miscoordination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 31 33 34 List of Figures 2.1 2.2 Matching pennies . . . . . . . . . . . . . . . . . . . . . . . . . . Classification of Stag-hunt game . . . . . . . . . . . . . . . . . . 8 11 3.1 General structure of Stag-hunt games . . . . . . . . . . . . . . . 14 4.1 Regret matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 5.1 5.2 5.3 Schmidt’s et al. Stag-hunt games . . . . . . . . . . . . . . . . . Battalio’s et al. Stag-hunt games . . . . . . . . . . . . . . . . . Dubois’s et al. Stag-hunt games . . . . . . . . . . . . . . . . . . 22 24 25 6.1 Experimental Stag-hunt games . . . . . . . . . . . . . . . . . . . 28 A.1 Instruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . II A.2 Translated instruction . . . . . . . . . . . . . . . . . . . . . . . III B.1 Game 1 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . IV B.2 Game 2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . V B.3 Game 3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . V C.1 Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . VI Bachelor Thesis Proposal Author Supervisor Proposed topic Michal Polena PhDr. Lubomı́r Cingl Does Payoff Dominance Matter? An Experiment. Topic characteristics Two selection criteria, a Risk dominance criterion and a Payoff dominance criterion, have been used for explaining outcomes in Staghunt games since they were introduced. Schmidt et al. (2003) find out that changes in Risk dominance parameter influence players’ choices but changes in Payoff dominance parameter has no impact on players’ choices. Schmidt et al. (2003)’s findings may hold true for relatively small changes in Payoff dominance parameter but the impact of large changes remain unanswered. The aim of this bachelor thesis is to examine whether large differences in Payoff dominance parameters have an impact on players’ decisions. An experiment with large changes in Payoff dominance parameter will be proposed and conducted in order to answer our research question. Outline 1. Introduction 2. Literature Review 3. Stag-hunt Game 4. Selection Criteria 5. Experimental Design 6. Results 7. Conclusion Bachelor Thesis Proposal xi Core bibliography 1. Battalio, R., L. Samuelson, & J.V. Huyck (2001): “Optimization incentives and coordination failure in laboratory stag hunt games.” Econometrica 69(3): pp. 749-764. 2. Carlsson, H. & E. V.Damme (1993): “ Global Games and Equilibrium Selection.” Econometrica 61 pp. 989-1018 3. Dubois, D., M. Willinger & P. Van Nguyen (2011): “Optimization incentive and relative riskiness in experimental stag-hunt games.” International Journal of Game Theory 41(2): pp. 369-380. 4. Feltovich, N., A. Iwasaki & S. H. Oda (2012): “Payoff levels, loss avoidance, and equilibrium selection in games with multiple equilibria: an experimental study.” Economic Inquiry 50(4): pp. 932-952. 5. Harsanyi, J. C. (1995): “A new theory of equilibrium selection for games with complete information.” Games and Economic Behavior 8(1): pp. 91-122. 6. Harsanyi, J. C., & R. Selten (1988): “A General Theory of Equilibrium Selection in Games.” Cambridge: The MIT Press. 7. Rasmusen, E. (2006): “Games and Information: An Introduction to Game Theory.” Malden: Blackwell Publishing, 4 edition. 8. Rydval, O. & A. Ortmann (2005): “Loss avoidance as selection principle: evidence from simple stag-hunt games.” Economics Letters 88(1): pp.101-107. 9. Schmidt, D., R. Shupp, J. M. Walker & E. Ostrom (2003): “Playing safe in coordination games. Games and Economic Behavior.” Games and Economic Behavior 42(2): pp. 281-299. 10. Skyrms, B. (2001): “The Stag Hunt.” Proceedings and Addresses of the American Phi losophical Association 75(2): pp. 31-41. Author Supervisor Chapter 1 Introduction A classic dilemma people are confronted with in their everyday lives is the decision between safe and risky, but potentially more rewarding, choices. Most of these decisions, such as international conflicts or campaign elections, can be made only once. In game theory, the dilemma can be modelled by Stag-hunt game, incorporating both opposing choices. Moreover, one-shot game setting can represent the decision occurring only once. Harsanyi & Selten (1988) introduced two selection criteria (Risk dominance and Payoff dominance criterion) helping players to solve the dilemma. There has not been a consensus on which of the criteria is more prominent in determining players’ decisions, until Schmidt et al. (2003) closely investigated the effect of parameters of given selection criteria. Schmidt et al. (2003) found out that players respond to changes in Risk dominance parameter, but do not respond to changes in Payoff dominance parameter. Nevertheless, Schmidt et al. (2003) did not control for other parameters of their games, such as Optimization premium and Relative riskiness. In addition, changes in Payoff dominance parameters in Schmidt et al. (2003)’s games might be considered too small to induce players to change their decisions. Experimental results of Battalio et al. (2001) and Dubois et al. (2011) suggest that sufficiently large differences in Payoff dominance parameter might explain their results. The purpose of this thesis is to examine whether large changes in Payoff dominance parameter might influence players’ decisions, keeping all other parameters equal. We proposed, therefore, three Stag-hunt games where we changed the magnitude of Payoff dominance parameter between Game 1 and Game 2. Moreover, we changed the magnitude of Relative riskiness parameter between Game 2 and Game 3 to investigate the predictive power of Relative 1. Introduction 2 riskiness. Our main finding is that large changes in Payoff dominance parameter do not influence players’ decisions. Thus, combining our findings with Schmidt et al. (2003)’s findings, we conclude that players do not respond to any changes in Payoff dominance parameter at all. Moreover, results of Game 2 and Game 3 show an insignificant effect of Relative riskiness parameter. Potential explanation are provided. The thesis is organised as follows. In Chapter 2, the terminology of game theory is defined. The Stag-hunt game is introduced and described in Chapter 3. Chapter 4 is dedicated to selection criteria and parameters used in Staghunt games. Literature review is presented in Chapter 5. Chapter 6 describes experimental procedure and results. Our main findings are summarized in Chapter 7. Chapter 2 Game Theory 2.1 A Brief Introduction to Game Theory Game theory is a technique used to analyze, model and solve a broad range of strategic situations in decision making. Different types of strategic situations, such as conflict situations or cooperative situations, may arise between individuals, groups or institutions every time and everywhere, when one’s payoff is not solely determined by one’s particular actions but is contingent on decisions made by other player or players. Therefore, one’s strategic plan of actions must take expectations of actions taken by others into account. The aim of game theory is to analyze given strategic situation, model the full range of outcomes based on one’s and others actions, and then provide one with optimal strategic plan. The optimal strategic plan is a set of one’s best-respond actions to others actions earning the highest possible payoff.1 Despite the fact that the notion theory is contained in game theory’s name, the game theory is not solely theoretical discipline. Game theory is a part of applied mathematics and has many useful applications in practice. The study of elections in political science, negotiations in diplomacy, evolutionary game theory in biology or multi-agent systems in computer science may serve as examples. The main application of game theory can be, however, found in economics. The economic phenomena, such as oligopolies, auctions, mechanism design and general equilibrium, are primarily analyzed by the use of game theory. 1 Carmichael (2005) 2. Game Theory 2.2 4 Short Historical Overview Origin of game theory can be traced back to the 17th century, where gambling was in the background of the development of mathematical probability. Pierre de Fermat and Blaise Pascal were the first pioneers in the mathematical probability at that time. Further development in mathematics was then used for studying and analyzing board games, such as chess or draughts. The very first discussion about game theory was found in the letter of James Waldegrave in 1713 (Bellhouse 2007). In this letter, an optimal strategy for solving a card game was described. The Researches into the Mathematical Principles of the Theory of Wealth was published by Antoine Augustin Cournot in 1838. Duopoly situations in the markets were studied in this research as well as a solution to these situations in the form of balanced strategies was provided (Romp 1997). The Researches into the Mathematical Principles of the Theory of Wealth is the first game theoretical study in economic science. The establishment of game theory as a new discipline; however, took place in the first half of the 20th century. The study ”Zur Theorie der Gessellschaftsspiele” (in English translated as ”On the Theory of Games of Strategy” (Tucker & Luce 1959) ) written by John von Neumann in 1928 is the foundation stone of game theory as a discipline. In this study, von Neumann proved a minimax theorem. The minimax theorem states that every finite, two-person, zero-sum game has a rational solution in the form of pure or mixed strategy(Poundstone 1992). The importance of the minimax theorem is demonstrated by von Neumann’s famous quote: ”As far as I can see, there could be no theory of games ... without that theorem ... I thought there was nothing worth publishing until the Minimax Theorem was proved”(Casti 1997). A groundbreaking book ”Theory of Games and Economic Behavior” was written by John von Neumann together with Oskar Morgenstern in 1944. The book presented applications of game theory in other disciplines, such as political science or diplomacy. This fact started an immense development and interest in game theory in the second half of the 20th century. Moreover, Theory of Games and Economic Behavior serves as a basis for current game theory. John Forbes Nash, Reinhard Selten and John Harsanyi were the most influential scholars in game theory in the second half of the 20th century. Nash proposed an equilibrium of non-cooperative games in which no player has any incentive to change his or her decision. This equilibrium is nowadays called a 2. Game Theory 5 Nash equilibrium. Selten improved the concept of Nash equilibrium with his solution principle of subgame perfect equilibria. Harsanyi created new concepts in game theory, such as complete information and Bayesian games. All three scholars received together the Nobel Prize in Economic Sciences 1994 for their contribution to game theory. In the 21st century, game theory is still developing, and its interdisciplinary applications are considerably expanding. Two Nobel Prizes in Economic Science have been awarded in this century. The first Nobel Prize was shared by Thomas Schelling and Robert Aumann in 2005. Schelling and Aumann received the Nobel Prize ”for having enhanced our understanding of conflict and cooperation through game-theory analysis”2 . Two years later, in 2007, the second Nobel Prize was given to Leonid Hurwicz, together with Eric S. Maskin and Roger B. Myerson. Hurwicz, Maskin and Myerson become Nobel laureates ”for having laid the foundations of mechanism design theory” 3 . 2.3 Game Theory Definitions In real life, the rules of a game must be set and described before playing a given game. This chapter, similarly to the real life approach, will define and describe rules of games in game theory. These rules will be essential for understanding the specific features of Stag-hunt game in the next chapter. The following definitions are based on textbooks of game theory (Gibbons (1992), Rasmusen (2006) and Carmichael (2005)). In the first instance, the components of the general game will be defined. Afterwards, normal-form representation of games will be introduced as a specific form of games. Normal-form games are frequently represented in matrixes (e.g. Figure 2.1). 2 Nobelprize. Retrieved from: sciences/laureates/2005/ 3 Nobelprize. Retrieved from: sciences/laureates/2007/ http://www.nobelprize.org/nobel prizes/economichttp://www.nobelprize.org/nobel prizes/economic- 2. Game Theory 6 Def.: Elements of a Game Essential components of a game are: a set of players P who make decisions an action aij which can be chosen by player i a set of all actions Ai = {aij } available to player i a player i ’s strategy si which is a plan what action(s) to pick out in a given moment of a game ; s−i represents strategies of all other players but player i a player i ’s payoff function πi (s1 , . . . , si , . . . , sn ) which shows payoffs de- pending on player i ’s and others actions Def.: Normal-Form Representation of Games Normal-form representation of a game is a description of a game in which each player simultaneously selects a strategy, and the resulting payoff is determined as an intersection of chosen strategies. Player i ’s pure and mixed strategies si determine probability with which a given action aij from the action set Ai is played. Def.: Pure Strategy Pure strategy si is a rule of player i that an action aij is played with probability one and any other actions from the actions set Ai with probability zero. Def.: Mixed Strategy Mixed strategy si is a rule of player i that each action aij from the action set Ai is played with certain probability pij . It holds that 0 ≤ pij ≤ 1 for j = 1, . . . , N and pi1 + . . . + piN = 1. The aim of mixed strategies is to choose such probabilities for given pure strategies in order to make other player indifferent between her choices. Making other player indifferent between her choices is the best strategy how to maximize own expected payoff. In addition, in some games, such as Matching pennies (see figure 2.1), deterministic pure strategies cannot be applied, and mixed strategies are used instead of pure ones. 2. Game Theory 7 Def.: Best Response Player i ’s strategy s∗i is called the best response to the strategies s−i selected by other players if this strategy earns her the greatest possible payoff. Mathematically, πi (s∗i , s−i ) ≥ πi (s0i , s−i ) ∀s0i 6= s∗i . If no other strategy is equally good, the best response strategy s∗i is strongly best. The best response strategy s∗i is weakly best if there is at least one equally good strategy. (Considering our mathematical entry, the definition of weakly best response is satisfied with equality.) Def.: Strictly Dominant Strategy The strategy s∗i is a called a dominant strategy if it yields higher utility than all other strategies in all possible states of the world. Formally, πi (s∗i , s−i ) > πi (s0i , s−i ) ∀s−i ∀s0i 6= s∗i . Weakly dominant strategy s∗i earns a player i at least as high payoff as any other strategies. Similarly to the strictly dominant strategy, πi (s∗i , s−1 ) ≥ πi (s0i , s−1 ) ∀s−i ∀s0i 6= s∗i . A player is advised by best response what to play given the other players’ actions. Regardless of the other player’ actions, a strictly dominant strategy recommends an optimal strategy to a player. Def.: Nash Equilibrium The mutual combination of players’ best responses (s∗1 , . . . , s∗n ) forms a Nash equilibrium. No player has an incentive to change his or her strategy in the Nash equilibrium. Formally, πi (s∗i , s∗−i ) ≥ πi (s0i , s∗−i ) ∀i ∀s0i . We can distinguish two types of Nash equilibria: Pure Strategy Nash Equilibrium and Mixed Strategy Nash Equilibrium. When we usually speak about the Nash equilibrium, we mean the Nash equilibrium in pure strategies. There are, however, some games, such as Matching pennies (see Figure 2.1), which do not have any Nash equilibrium in pure strategies. Nevertheless, there must be 2. Game Theory 8 an equilibrium in every finite game (Nash (1950)). Therefore, games without the Nash equilibrium in pure strategies have the Nash equilibrium in mixed strategies. Moreover, there are games with multiple Nash equilibria that have Nash equilibrium in pure strategies as well as in mixed strategies. An example of a game with both types of equilibria is a Stag-hunt game. Figure 2.1: Matching pennies . Def.: Risk Dominant Nash Equilibrium Risk Dominant Nash Equilibrium is formed as an intersection of players’ risk dominant strategies4 . Def.: Payoff Dominant Nash Equilibrium In a game with multiple Nash equlibria which are Pareto-rankable 5 , the Payoff dominant Nash Equlibrium is the one which earns a higher payoff to each player than the other equilibria. In games with multiple Nash equilibria, such as Stag-hunt games, where one equilibrium is Risk dominant Nash equilibrium and the other one is Payoff dominant Nash equilibrium, it is not sure which equilibrium will be played. Therefore, there are selection criteria and parameters of a game that may predict which Nash equilibrium will be played. 4 Strategies selected by Risk dominance criterion are called risk dominant strategies (See 4.2 Risk Dominance Criterion) 5 The term, Pareto-rankable equilibria, was firstly used by Thomas Schelling in his book The Strategy of Conflict. This term is used for payoff distinguishable equilibria in games with multiple Nash equilibria. 2. Game Theory 2.4 9 Classification of Games Games can be classified according to many different criteria. This section is particularly concerned with classification criteria which are relevant to the Staghunt game. The classification of games will be helpful for understanding and describing the Stag-hunt game in the next chapter. According to the Volek (2010), games can be classified with following classification criteria: Number of Players Games can be classified according to the number of players who play a given game. We distinguish games with 1, 2 or n players. Number of Strategies There is either a finite number of strategies or infinite number of strategies in an action set Ai . Degree of Allowed Communication In communication setting, players are allowed to communicate among themselves. Therefore, players might be likely to reach an agreement on given actions and to make a binding commitment. Communication among players can be further discriminated between one-way and two-way communication. In one-way communication, one player can communicate her intention to other players. Other players, however, are not allowed to respond. Thus, the player will not receive any feedback from other players. On the contrary, in two-way communication players can respond to each other. As opposed to communication setting, players are not allowed to communicate among themselves in non-communication games. Hence, players cannot make any binding commitments. Character of Payoffs Games can be classified according to their sum of all players’ payoffs. We distinguish constant sum games (antagonistic games) and non-constant sum games (non-antagonistic games). In constant sum games, a gain of one player entails a loss of other player because the sum of all players’ payoffs is a given constant 6 . Consequently, the actions of players are in 6 The most well-known type of constant sum game is zero-sum game. Figure 2.1: Matching pennies is an example of zero-sum game. 2. Game Theory 10 conflict of interest. In non-constant sum games, the player’s actions are not in conflict of interest and they may cooperate. Amount of Information Players can either have complete or incomplete information in a game. Complete information means that the structure of the game and the player’s payoff function is common knowledge among all the players (Gibbons (1992)). On the contrary, incomplete information means that either information about the structure of game or information about the player’s payoff function is not known by players. Number of Periods We distinguish one-shot games and repeated games. One-shot games are played only once. Repeated games are played many times. Repeated game can be played with fixed number of periods or with some random number of periods. Random number of periods may be determined with a device which at the end of every period with some probability decides whether the game will continue or not. Moreover, repeated games are played with fixed or random matching. Fixed matching means that player is randomly matched with another player in the first round but she will stay fixed to that player in the following games. In contrast to the fixed matching, in random matching players are matched with another randomly chosen player in every single game. Symmetry Games can be either symmetric or non-symmetric. A symmetric game means that the choice of a player’s position, whether you are player 1 or player 2, does not change your payoff function. In other words, given other player’s choice, you will earn the same amount of money irrespective of your position as player 1 or player 2. On the contrary, your payoff function in non-symmetric games is dependent on your player’s position. According to the Bruns (2010), there are only 12 symmetric games out of all 144 types of distinct 2x2 games. 2. Game Theory 11 The following figure is based on decision-making situation scheme from Volek (2010). The red circles denote classification features of Stag-hunt games. Figure 2.2: Classification of Stag-hunt game Source: Author’s translation of decision-making situation scheme from Volek (2010) . Chapter 3 Stag-hunt Game 3.1 An Introduction to Stag-hunt games ”If it was a matter of hunting a deer, everyone well realized that he must remain faithfully at his post; but if a hare happened to pass within the reach of one of them, we cannot doubt that he would have gone off in pursuit of it without scruple and, having caught his own prey, he would have cared very little about having caused his companions to lose theirs.” { Jean-Jacques Rousseau } According to the Skyrms (2001), the Stag-hunt game’s structure as well as the name came from the above mentioned story. The story about hunting a deer comes from Discourse on Inequality, book written by Jean-Jacques Rousseau in 1754. Moreover, Skyrms (2001) claims that examples of Stag-hunt games can be found in many philosophical works. Two men pulling at the oars of a boat from A Treatise of Human Nature written by David Hume is one of the examples. The Stag-hunt game is considered to be one of the most important games in game theory because it includes the features and dilemmas of social contracts. The dilemma between the safe choice and the risky choice with a possible payofffavourable outcome might be found in the business as well as in the real life. A business example of Stag-hunt game might be two telecommunication firms offering similar LTE network access. Advertising advantages of LTE 1 over 3G will raise the consumer’s demand for LTE. If both firms advertise advantages 1 LTE is an abbreviation for Long-Term Evolution. LTE and 3G are communication networks for high-speed data. 3. Stag-hunt Game 13 of LTE, they will both greatly benefit from it. If one of the firms does not advertise at all, the firm will earn more than the advertising firm because the advertising firm has some costs. But the non-advertising firm will benefit less than in the case when both firms advertise. The real life example of Stag-hunt game might be the decision of young graduates in Mississippi whether to stay or not in their country. 2 In game theory, the Stag-hunt game is also known as coordination game, security dilemma, trust dilemma or assurance game. We would, however, argue that the name coordination game is an inaccurate appellation for Stag-hunt game. Furthermore, we claim that the Stag-hunt game should not consider to be a coordination game at all. According to the Rasmusen (2006), coordination games are games which share the common feature that the players need to coordinate on one of the multiple Nash equilibria. In Stag-hunt games, players wish to coordinate solely on the Payoff dominant equilibrium. If players are in doubt about coordination on the Payoff dominant equilibrium, they can choose the risk dominant choice. If players choose the risk dominant choice, they do not need other player to coordinate on the risk dominant Nash equilibrium. On the contrary, every player choosing risk dominant choice will be at least weakly better off if other player selects the payoff dominant choice. This obviously stands in contradiction with the definition of coordination games. 3.2 Stag-hunt Game in the Perspective of Game Theory There is a large number of different types of games in game theory. For the purpose of this Bachelor thesis, it is, therefore, necessary to clearly classify the Stag-hunt game with its characteristics . From now on, we will solely consider games with following characteristics: Number of Players All games are played by two players. Number of Strategies Every player has two pure strategies in her action set Ai . 2 RethinkMississippi. Retrieved from: http://www.rethinkms.org/2014/01/15/mississippianshunt-deer-rabbits-brain-drain-explained-game-theory/ 3. Stag-hunt Game 14 Degree of Communication Only games without communication are considered. Character of Payoffs Stag hunt games are non-constant sum games. Amount of Information Every player has complete information about the game. Number of Periods One-shot games and first rounds of games with random matching are solely considered. Schmidt et al. (2003) found that there are no significant differences between one-shot games and first rounds of games with random matching. Schmidt et al. (2003)’s claim is further supported by Gallice (2006) and Guyer & Rapoport (1972). It is argued that games with random matching are similar to the one-shot games. Players cannot play strategically with respect to the next games as they are rematched in each period. Moreover, players cannot experience the learning effect because only the first rounds are considered, Symmetry Every Stag-hunt game is symmetric. 3.3 Structure of Stag-hunt Game The general structure of Stag-hunt game is depicted in normal-form representation of games in Figure 3.1 . Figure 3.1: General structure of Stag-hunt games 3. Stag-hunt Game 15 Following the definitions from section 2.3, the general elements of a game are in our case two players (player 1 and player 2). Each of them has two pure strategies in her actions set. Player 1 has pure strategies U and V in her action set A1 . Player 2 has pure strategies L and R in her action set A2 . The last element of a game is a payoff function represented by the lower and upper case letters. The upper case letters A, B, C and D are payoffs of player 1. The lower case letters a, b, c and d are payoffs of player 2. Player’s payoffs are determined as the intersection of their choices. For example, if player 1 plays U and player 2 plays R, then the intersection of players’ choices is the upper-right matrix with payoff B for player 1 and payoff c for player 2. The specific structure of Stag-hunt game is that A > C, D > B, D > A and B ≥ A. Some games are called Stag-hunt games even though they do not satisfy B ≥ A. The requirement for B ≥ A, however, is related to the story about hunting a stag (in Figure 3.1 represented by choices V and R). Less or at least the same effort is exerted for hunting a hare (in Figure 3.1 represented by choices U and L) when other player hunts the stag than in situation when other player hunts the hare too. It should be easier to catch the hare when other player hunts the stag. Therefore, the payoffs in Stag-hunt game should satisfy B ≥ A. Preceding structural statements hold for lower case letters in the same way as for upper case letters because the Stag-hunt game is symmetric. The best response for player 1 is to play U (V ) when player 2 plays L(R) because the payoff A(D) is higher than C(B). The same argument holds for player 2 with his or her payoffs in lower case letters. The Stag-hunt game has neither strictly nor weakly dominant strategies. Both players have the same mutual best responses to other player’s decisions. Therefore, there are two Nash equilibria in pure strategies. The first Nash equilibrium in pure strategies is {U ; L} and the second one is {V ; R}. According to the Nash equilibrium definition, neither player has an incentive to deviate from her best response. It holds because A > C and a > c for the first equilibrium {U ; L} as well as D > B and d > b for the second equilibrium {V ; R}. There are two Nash equlibria in Stag-hunt games which yield different payoffs. Thus, the equilibria can be defined as Pareto-rankable equilibria. The equilibrium {V ; R} is Payoff dominant Nash equilibrium because its payoffs D for player 1 and d for player 2 are strictly higher than payoffs A and a from equilibrium {U ; V }. The inferior Nash equilibrium {U ; V } is Risk dominant Nash equilibrium. 3. Stag-hunt Game 16 The enumeration of Nash equilibria is, however, not complete yet because there is a remaining equilibrium. It is a Nash equilibrium in mixed strategies. To maximize one’s own expected payoff in mixed strategies, one should play her strategies with such probabilities that the other player’s expected payoffs are the same regardless of her choice. Therefore, the other player is indifferent between her choices. In other words; considering the general structure of Staghunt game, if player 2 wants to make player 1 indifferent between her choice U and V , then player 2 must assign probabilities p and 1 − p to her choices L and R in order to do so. Player 1’s expected payoffs are Ap+B(1−p) by playing U and Cp+D(1−p) by playing V . The solution of equation where player 1’s expected payoffs are equal, formally Ap + B(1 − p) = Cp + D(1 − p), provides player 2 with optimal probability p for mixing her choices. The optimal probability p is p= A−C . A−C +D−B A−C Player 2 should play strategy L with probability p = A−C+D−B and strateD−B gy R with 1 − p = A−C+D−B to make player 1 indifferent between her choices. Player 1 can make player 2 indifferent in exactly the same way by assigning probabilities q and (1 − q) to her choices U and V . As the Stag-hunt games are symmetric, it holds that p = q. The Nash equilibrium in mixed strategies can be written as {(U, p = D−B )}. A−C+D−B A−C ; V, 1 A−C+D−B −p = D−B ); (L, q A−C+D−B = A−C ; R, 1 A−C+D−B −q = Player’s choices U and L are called risk dominant strategies because payoffs following from playing U or L are being considered safe in comparison with payoffs achieved by playing choices D or R. Choices D and R are called pareto dominant strategies because in the case of successful coordination payoff dominant Nash equilibrium is achieved. Therefore, the trade-off between risk dominant and pareto dominant strategies constitutes an interesting dilemma in Stag-hunt games. Chapter 4 Selection Criteria & Parameters Selection criteria and particular parameters of Stag-hunt games will be introduced and defined in this chapter. Selection criteria provide us with recommendations which strategy to play in given games. Parameters give us important information about the payoff structure of games, with suggestion what one should choose. Both, selection criteria and parameters, are useful for decisionmaking in the presence of multiple Nash equilibria with the trade-off between risk-safe strategy and risky but pareto efficient strategy. The algebraic notion from section 3.3 Structure of Stag-hunt Game is used for explaining the selection criteria. 4.1 Payoff Dominance Criterion ”Rational individuals will cooperate in pursuing their common interests if the conditions permit them to do so . . . ” {Harsanyi & Selten (1988)} The above stated quote is the core idea of Payoff dominance criterion. Harsanyi & Selten (1988) argue in their famous book A General Theory of Equilibrium Selection in Games that rational players should coordinate on the Payoff dominant Nash equilibrium. In other words, player 1 and player 2 would play strategies V and R leading to the most efficient equilibrium with payoffs D and d. Hence, Harsanyi & Selten (1988) prefer the Payoff dominance criterion to the Risk dominance criterion (described in the following section). Choices selected by Payoff dominance criterion are called pareto dominant strategies. Schmidt et al. (2003) introduced Payoff dominance parameter P as an efficiency loss measurement. Parameter P measures the percentage efficiency 4. Selection Criteria & Parameters 18 loss incurred by successful coordination on the inferior equilibrium in comparison with successful coordination on the Payoff dominant equilibrium. The parameter P is defined as P = p(D, R) − p(U, L) D−A d−a = = p(D, R) D d where p(D, R) means payoffs from choice D played by player 1 and from choice R played by player 2. 4.2 Risk Dominance Criterion The Risk dominance criterion is the second selection criterion proposed by Harsanyi & Selten (1988). This selection criterion is based on the comparison of Nash products of given Nash equilibria. Nash product is a payoff loss of both players incurred by their deviation from Nash equilibrium. In our case, the Nash product of Nash equilibrium {U ; L} is (A − C)(a − c) and the Nash product of Nash equilibrium {V ; R} is (D − B)(d − b). Harsanyi & Selten (1988) claim that the equilibrium with the higher Nash product should be chosen. Strategies chosen by Risk dominance criterion are called risk dominant strategies. In Stag-hunt games, the Payoff dominance criterion and the Risk dominance criterion have often opposite recommendations. Whilst strategies V and R are recommended by the Payoff dominance criterion, the Risk dominance criterion may recommend strategies U and L. The Risk dominance criterion, however, may recommend payoff dominant strategies V and R too. For our purposes, we will use slightly modified version of Risk Dominance Criterion. The modified version of Risk Dominance Criterion measured by parameter R was proposed by Selten (1995). The parameter R is defined as R= A−C a−c p(U, L) − p(V, L) = = p(V, R) − p(U, R) D−B d−b where p(U, L) means payoffs (A and a) from intersection of choice U played by player 1 and from choice L played by player 2. Recommendations given by the modified vision are the same as recommendations by Nash products. If parameter R is positive, the strategies, U and L leading to the equilibrium {U ; L} are recommended. If parameter R is equal to zero, the mixed strategies are the risk dominant strategies. Strategies V and R leading to the equilibrium {V ; R} are recommended if parameter R is negative. 4. Selection Criteria & Parameters 4.3 19 Optimization Premium As opposed to Payoff and Risk dominance criteria, Optimization premium is rather parameter than selection criterion. The Optimization premium was invented by Battalio et al. (2001). The Optimization premium parameter, for our purposes labelled OP, is defined as OP = [p(U, L)−p(V, L)]+[p(V, R)−p(U, R)] = [A−C]+[D−B] = [a−c]+[b−d]. It is simply the absolute sum of deviation losses from single Nash equilibria added up together. Battalio et al. (2001) argue that ceteris paribus 1 the higher the Optimization premium parameter is, the more likely players will choose the risk dominant strategies.2 4.4 Relative Riskiness Similarly to the Optimization premium, the Relative riskiness is parameter of the payoff structure of Stag-hunt games. The Relative riskiness was proposed by Dubois et al. (2011). The Relative riskiness parameter is labelled RR, and is defined as B−A b−a p(U, R) − p(U, L) = = . RR = p(V, R) − p(V, L) D−C d−c Dubois et al. (2011) argue that ceteris paribus the closer is the Relative riskiness to zero, the safer the risk dominant strategy seems to be, and players are more likely to choose the risk dominant strategy. On the contrary, if the Relative riskiness is close to one, the pareto and risk dominant strategies embrace the same amount of risk. There is, however, evident shortcoming in Relative riskiness parameter. Dubois et al. (2011) assume that payoffs difference p(U, R) − p(U, L) should not be zero. Otherwise, the value of p(U, R) and p(U, L) would not have any effect on Relative riskiness. This strong assumption is not always achievable in Stag-hunt games, which consequently lowers applicability of Relative riskiness. 1 Ceteris paribus means with other conditions remaining the same or other things being equal. OxfordDictionaries. Retrieved from: http://www.oxforddictionaries.com/definition/english/ ceteris-paribus 2 From now on, risk dominant strategies refer to players’ strategies leading to the inferior Nash equilibrium. Using labelling from section 3.3, players’ strategies U and L are considered to be risk dominant strategies. Reader will be warned when risk dominant strategies are meant to be pareto dominant strategies. 4. Selection Criteria & Parameters 4.5 20 Mixed Strategies & Mixed Minimax Regret Mixed strategies are considered to be selection criterion (Gallice (2006)) as well as particular parameters of Stag-hunt games (Battalio et al. (2001)). The core idea behind mixed strategies was already explained in Chapter 3. The inference from Chapter 3 was that players should play strategies U and L A−C a−c with probability pM S = A−C+D−B = a−c+d−b to make other player indifferent between her choices. Mixed Minimax regret is selection criterion proposed by Gallice (2006). The computation of the mixed Minimax regret is made in two steps. Firstly, the common matrix must be altered to the regret matrix. The matrix from Figure 3.1 was altered into to the regret matrix in Figure 4.1. Gallice (2006) defines particular payoffs in the regret matrix as regret. The regret is nonnegative difference between the received payoff and the payoff which would be received if the best response was chosen. For example, the payoffs zero for both players as intersection of their choices U and L means that originally received payoffs are equal to the payoffs from best responses. Thus, both players played optimally, and their regret is equal to zero. Figure 4.1: Regret matrix Secondly, the same approach as in Mixed strategies is applied to the regret matrix. Player 2 properly assigned probabilities to her strategies if player 1’s expected regrets are equal: p0 + (1 − p)(D − B) = p(A − C) + (1 − p)0. Solution of the preceding equation is mixed Minimax regret probability: pM R = D−B d−b = . A−C +D−B a−c+d−b As the regret matrix is symmetric, both players have the same probability pM R for playing their strategies U and L. One interesting fact is that probabilities from Mixed strategies and mixed Minimax regret are complementary. In other words, they add up to one pM S + pM R = 1 (Gallice (2007)). Chapter 5 Literature Survey There has not been a consensus on whether players follow Payoff dominance or Risk dominance criterion since both criteria were introduced in A General Theory of Equilibria. Harsanyi & Selten (1988), Schelling (1960) and Anderlini (1999)) believe that player should behave according to the Payoff dominance criterion and choose the pareto dominant strategy. On the other hand, Carlsson & Damme (1993), Huyck et al. (1990) and Harsanyi (1995) prefer the Risk dominance criterion. Although no final conclusion has been drawn yet, there is prevailing agreement that a Risk dominance criterion is more salient for outcome prediction (Gallice (2006)). The prevalence of Risk dominance salience is supported by experimental evidence (Cooper et al. (1992), Straub (1995), Huyck et al. (1990)) where subjects chose risk dominant strategies more often than payoff dominant strategies. Keser & Vogt (2000) proposed and conducted an experiment with a game in which payoff dominant strategy and risk dominant strategy were identical. In other words, both selection criteria recommended the same strategy. Therefore, there should have been a clear outcome predicted by selection criteria. More than 40% of subjects, however, chose a strategy that was payoff dominated as well as risk dominated by another strategy. It seems that subjects do not exclusively decide according to the given selection criteria. Parameters of selection criteria, such as parameter P of Payoff dominance criterion and parameter R of Risk dominance criterion, should be rather considered. Parameters include important information about the magnitude of particular selection criterion. The reason why more than 40% of subjects chose payoff and risk dominated strategy could be explained by the magnitude of parameter P and R. Moreover, other parameters, such as OP parameter or RR parameter, could play a 5. Literature Survey 22 role. Schmidt et al. (2003) closer examined the importance and the predicted power of parameters P and R. Schmidt et al. (2003) proposed four different Stag-hunt games (depicted in Figure 5.1)1 . Authors argue that the differences in decisions between games G2 and G3 as well as the differences between games G1 and G4 are solely attributed to the changes in parameter R because the parameter P is held constant. Similarly, the differences in decisions between games G2 and G4 as well as between games G1 and G3 are attributed to the changes in parameter P while R is held constant. Schmidt et al. (2003) conducted their experiments in three different settings: one-shot games, repeated games with random matching and repeated games with fixed matching. Figure 5.1: Schmidt’s et al. Stag-hunt games Legend: R is Risk dominance parameter. P is Payoff dominance level parameter. OP is Optimization premium parameter. RR is Relative riskiness parameter. MS stands for Mixed strategies. The results from one-shot games and repeated games with random matching revealed that players respond to the changes in parameter R while parameter 1 The Optimization premium parameter (OP) and the Relative riskiness parameter (RR) were not included in Schmidt et al. (2003) study. These parameters are added for purposes of this bachelor thesis. 5. Literature Survey 23 P was kept constant. In one shot games2 , 40% of players played risk dominant strategies in games G1 and G3. After increasing the value of parameter R from R = log1 to R = log3, 60% of players in game G2 and 58% of players in game G4 played the risk dominant strategies. In other words, the change in parameter R accounts for roughly 19% change in player’s decisions which is a significant difference according to Schmidt et al. (2003). On the contrary, the experimental results did not reveal any significant differences between player’s choices when parameter P was changed, and parameter R was kept constant. In games G1 and G3, 40% of players chose the risk dominant strategy with parameter P = 0.2 , and the same proportion of players chose the risk dominant strategy after parameter P was increased to P = 0.4. Similar results were obtained in games G2 and G4, where 60% of players chose the risk dominant strategy before and 58% of players after the parameter P was increased. Considering obtained results, Schmidt et al. (2003) concluded that players do not respond to the changes in parameter P . The results of Schmidt et al. (2003), however, do not seem to be in line with experimental evidence from Battalio et al. (2001) and Dubois et al. (2011). Moreover, combined experimental evidence from Feltovich et al. (2012) and Straub (1995) contradicts the Schmidt et al. (2003)’s results as well. Battalio et al. (2001) proposed and conducted an experiment on three different repeated Stag-hunt games with random matching. (see Figure 5.2)3 . They wanted to prove the influence of Optimization premium on player’s decisions. Battalio et al. (2001)’s results are in line with their hypotheses about Optimization premium, but they considered neither parameter P nor parameter RR in their experiment. For purposes of this thesis, only results from first periods are reported as a proxy for one-shot games. Moreover, results are interpreted with the use of the same parameters as Schmidt et al. (2003) used in their study. All three games had the same magnitude of parameter R = log4 and increasing magnitude of parameter P game by game. According to the inferences from Schmidt et al. (2003), there should not have been any significant difference between games because parameter R was kept constant. There was no significant difference 2 We decided to demonstrate the Schmidt’s main findings on one-shot games. The inferences from results of repeated games with random matching are, however, the same as in the one-shot game setting. 3 In line with the arguments stated in section 3.3 Structure of Stag-hunt Games, we do not consider game 2R to be a proper Stag-hunt game because it does not satisfy A ≥ B. Moreover, the game 2R can be called coordination game because both players have an incentive to coordinate on one of the Nash equlibria. 5. Literature Survey 24 Figure 5.2: Battalio’s et al. Stag-hunt games between games R and 0.6R. Nevertheless, there was a significant difference between games 2R and R. Furthermore, the difference between games 2R and 0,6R was not significant, but the trend that with increasing parameter P players will be more likely to play payoff dominant strategies was obvious in this case. It seems that players might not react to small changes in parameter P . In Schmidt et al. (2003), the parameter P was changed from P = 0.2 to P = 0.4. Thus, the absolute change in parameter P was only 0.2, which may not be sufficient cause for players to change their decisions. If parameter P was increased from P = 0.11 to P = 0.55 or P = 0.733, the significant difference or at least the trend was detected. It is necessary to point out that other parameters, such as Optimization premium parameter (OP ) or Relative riskiness (RR), might have played a role in determination of Schmidt et al. (2003)’s results. This fact is another limitation of Schmidt et al. (2003) results because they did not control for other parameters than Risk dominance parameter, Payoff dominance parameter and Mixed strategies4 . Dubois et al. (2011) do not agree with the Battalio et al. (2001)’s inferences about Optimization premium. Dubois et al. (2011) argue that Battalio et al. (2001) unconsciously changed Relative riskiness parameter RR while changing their Optimization premium parameter OP . Dubois et al. (2011) proposed 4 It is not necessary to control for parameter pM S of Mixed strategies because the probability pM S is related to the the parameter R of Risk dominance. Thus, keeping the parameter R constant in different games means that parameter pM S remains constant too. 5. Literature Survey 25 and conducted an experiment on three repeated Stag-hunt games with random matching (depicted in Figure 5.3), in which either parameter RR or parameter OP was changed. Dubois et al. (2011)’s results reject conclusions from Battalio et al. (2001). Following the inferences from Schmidt et al. (2003), Dubois et al. (2011) did not control for Payoff dominance parameter P . The results of Dubois et al. (2011) could be, however, explained by use of parameter P . Considering the first rounds of games, there is an obvious trend in results that with an increase in parameter P players are more likely to choose the payoff dominant strategies. Even though there is the trend in player’s choices between Games 2 and 3, the difference between these games is not significant. This results is similar to the results of Schmidt et al. (2003) and Battalio et al. (2001), because the difference in Payoff dominance parameter level P is only 0.166 (P3 − P2 = 0.366 − 0.2). In the same way as in Battalio et al. (2001)’s experiment, we argue that this difference in parameter P is not sufficient for players to change their choices. On the other hand, there are significant differences in players’ choices between Game 1 and Game 2 as well as between Game 1 and Game 3. These differences in players’ choices can be explained by sufficient differences between Payoff dominance parameters. The difference between parameters P is 0.533 (P1 − P2 = 0.733 − 0.2) in Games 1 and 2, and 0.367 (P1 − P3 = 0.733 − 0.366) in Games 1 and 3. It seems that players are more likely to choose Payoff dominant strategies in games with sufficiently large Payoff dominance parameter P . Figure 5.3: Dubois’s et al. Stag-hunt games Chapter 6 Experiment Design The main finding of Schmidt et al. (2003) is that Payoff dominance parameter P does not influence players in selecting their choices in one-shot games as well as in first rounds of repeated games. The experimental results from Battalio et al. (2001) and Dubois et al. (2011) indicate that the parameter P might play a role in determining the outcome. Battalio et al. (2001) explain their experimental results with use of Optimization premium parameter OP . Dubois et al. (2011) reject the Battalio et al. (2001)’s results because Battalio et al. (2001) do not control for the Relative riskiness parameter RR in their experiments. According to the Dubois et al. (2011), the experimental results of Battalio et al. (2001) and Dubois et al. (2011) can be explained by use of parameter RR. Neither Battalio et al. (2001) nor Dubois et al. (2011), however, control for the Payoff dominance parameter P in their experiments. The aim of this Bachelor thesis is to closely examine whether the sufficiently large changes in Payoff dominance parameter P influences players’ choices in one-shot games. We question the main findings of Schmidt et al. (2003) because of two reasons. The first reason is that Schmidt et al. (2003) do not control for Optimization premium parameter OP , and the Relative riskiness parameter RR is not applicable to the Schmidt et al. (2003)’s games. The second reason is that the difference in parameters P between given games is only 0.2. We argue that the difference 0.2 in parameters P is not large enough to induce players to change their choices. Therefore, the results of Schmidt et al. (2003) may be only true for small differences in parameters P between given games. Moreover, the evidence from Battalio et al. (2001) and Dubois et al. (2011) might support our claim. The differences in players’ choices were not significant in games with small differences in parameters P . But there were significant differences 6. Experiment Design 27 in players’ choices for large differences in parameters P . The results of Battalio et al. (2001) cannot be taken as evidence that Payoff dominance parameter P matters because Battalio et al. (2001) does not control for parameter P as well as Relative riskiness parameter RR. With the change of Optimization premium parameter OP , Battalio et al. (2001) change unconsciously the values of parameters P and RR as well. Therefore, it is impossible to analyse the influence of particular parameters. Similarly to the Battalio et al. (2001)’s experiment, Dubois et al. (2011) do not control for parameter P . Thus, Dubois et al. (2011) always change at least two parameters OP , RR or P at the same time which makes it unfeasible to analyze effects of particular parameters. One could argue that we do not have to control for Optimization premium parameter OP because its significance is rejected. We agree with this argument. Nevertheless, even after omitting the influence of Optimization premium parameter OP , the Dubois et al. (2011)’s results cannot be taken as evidence for our hypothesis that parameter P matters. In Dubois et al. (2011)’s Stag-hunt Game 2 and Game 3, the relative riskiness parameter RR is kept constant but the difference between parameters P (0.366 − 0.2 = 0.16) in these games is even smaller than in Schmidt et al. (2003)’s experiment. Thus, these results are not taken into account because the difference in parameters P is not large enough. The differences in parameters P are considered to be large enough between Game 1 and Game 2 or between Game 1 and Game 3. But the Relative riskiness parameter RR is different in Game 1 in contrast to Game 2 and Game 3. Therefore, it is unclear whether parameter RR or parameter P influences players’ decision in this case. We proposed three Stag-hunt games (depicted in Figure 6.1) to closely examine whether the Payoff dominance parameter P does matter or does not, keeping all other parameters equal. All three games have the same values of Risk dominance parameter R, Optimization premium parameter OP and probability Mixed strategy pM S . The differences between games are in use and magnitude of parameters P and RR. Parameter P is 0.1 in Game 1, and 0.5 in Game 2 and Game 3. The Relative riskiness parameter RR is 41 in Game 1 and Game 2, and 12 in Game 3. The differences in players’ choices between Game 1 and Game 2 are solely attributable to changes in parameters P because parameter RR is held constant. The differences in players’ choices between Game 2 and Game 3 are exclusively caused by changes in parameter RR. 6. Experiment Design 28 Figure 6.1: Experimental Stag-hunt games Hypothesis 1: Significantly more payoff dominant strategies will be chosen in Game 2 than in Game 1. We hypothesize that players will be more likely to choose payoff dominant strategies in Game 2 than in Game 1 because of the changes in parameter P . The difference in parameters P between Game 1 and Game 2 is 0.4 (P2 − P1 = 0.5 − 0.1). We conjecture that the difference in parameters P is sufficiently large to induce players to play more often payoff dominant strategies. Hypothesis 2: No significant differences in players’ choices will be detected between Game 2 and Game 3. According to Dubois et al. (2011), players should be more likely to play risk dominant strategies in Game 2 than in Game 3 because of the Relative riskiness parameter RR. Dubois et al. (2011) argue that the closer is the parameter RR to zero the more likely are players to choose risk dominant strategies. We challenge the importance of parameter RR because of its limitations in applicability. Moreover, we conjecture that the results of Dubois et al. (2011) and Battalio et al. (2001) experiments can be rather explained by Payoff dominance parameter P than by Relative riskiness parameter RR. Therefore, we hypothesize that there will be no significant differences between Game 2 and Game 3 because the value of parameter P is the same for both games. Chapter 7 Experiment 7.1 Experiment Procedure The experiment was conducted in the Laboratory of Experimental Economics (LEE) at the University of Economics in Prague at the beginning of April 2014. Participants were recruited through ORSEE (Greiner (2004)). Experimental software z-Tree (Fischbacher (2007)) was used for programming and carrying out the experiment. In total, three sessions with 68 participants were conducted. The experiment was carried out in Czech. At the beginning of the experiment, printed instructions were given to participants. Original Czech instruction with their translation into English are enclosed in Appendix A. Each participant played all three one-shot games in three rounds. Every participant was matched to another participant after each round. Moreover, participants did not receive any feedback after each round. Thus, rounds were mutually independent. To avoid the order effect, there were six different types of order 1 . Only 20 participants out of 24 expected participants came to the first session. Therefore, the last type of order (G3 → G2 → G1) was not played in the first session. In the remaining two sessions, all six types of order were played by 24 participants in each session. At the end of experiment, all payoffs from previous rounds were displayed to participants. One of rounds was randomly chosen, and players were paid out according to their payoffs from chosen round. Experimental currency units obtained in Stag-hunt games were converted to the Czech currency in exchange rate one to one. Participants earned 43.60CZK(' $2.18) on average. Approximately, two pieces of bread 1 The types of orders were: G1 → G2 → G3 ; G1 → G3 → G2 ; G2 → G1 → G3 ; G2 → G3 → G1 ; G3 → G1 → G2 ; G3 → G2 → G1. 7. Experiment 30 can be bought for an average payment. The experiment was part of another experiment, otherwise, the payoffs would be too small for participants. The experiment was generously funded from the resources of LEE. Participants were asked to fill out a questionnaire about their demographic data. The following Table 7.1 summarizes demographic data about participants. Table 7.1: Demographic Data of Participants Session 1st Number of participants - Man - Woman Average age Degree* Field of Study 20 13 7 22.5 8 2nd Total 3rd 24 24 21 18 3 6 22.5 21.9 8 7 Economics or Business 68 52 16 22.3 23 * Degree involves partcipants with bachelor as well as with master degree. 7.2 Hypothesis Testing Result 1: There is no significant difference in players’ choices between Game 1 and Game 2. The experimental results do not support our first hypothesis that significantly more payoff dominant strategies (B) will be chosen in Game 2 than in Game 1. The results from the experiment are summarized in the Table 7.2. According to the Chi-Square test, the p-value is 0.3. Therefore, the difference is not significant at p < 0.05. We infer from our Results 1 that large differences in Payoff dominance parameter P do not influence players’ choices. Combining our results with existing literature about Payoff dominance parameter P , we conclude that Payoff dominance parameter P does not play any role in determining the outcome. It seems that players do not solely respond to the Nash equilibria in Stag-hunt games as expected by Payoff dominance parameter P . There are not significantly more payoff dominant choices in Game 2 than in Game 1 as we hypothesized. But not even a trend towards more payoff dominant choices in Game 2 is observable. On the contrary, there are more 7. Experiment 31 payoff dominant choices in Game 1 than in Game 2. We conjecture that the slightly higher number of payoff dominant choices in Game 1 in comparison with Game 2 is caused by random noise. In order to avoid some discrepancy in results interpretation, we discuss two other explanations in following paragraphs. Table 7.2: Experimental results Game 1 Game 2 Game 3 Decision A 27 33 27 % of A Decisions 39.7% 48.5% 39.7% Decision B 41 35 41 % of B Decisions 60.3% 51.5% 60.3% The presence of zero payoff in matrix cells {B; A} for player 1 and {A, B} for player 2 in Game 2 may be seen as another explanation for the higher number of payoff dominant choices in Game 1 than in Game 2. This argument would mean that players are afraid to play payoff dominant strategy B in Game 2 because it could result in zero payoff. In other words, players would be considered to be risk averse and choose, therefore, risk dominant choice A. Nevertheless, Neumann & Vogt (2009) argue that risk averse players are not more likely to choose risk dominant choice in Stag-hunt games. Thus, the presence of zero payoff should not induce players to play the risk dominant choice A. Considering Figure 7 with proposed Stag-hunt games, one might argue that the substantial Payoff level difference between Game 1 and Game 2 could play a role in determining the outcome. This argument may be refuted by Rydval & Ortmann (2005). Rydval & Ortmann (2005) argue that Payoff levels in Stag-hunt games do not play a role if the payoffs are non-negative. Moreover, the result from Game 3 supports the Rydval & Ortmann (2005)’s claim. The number of payoff dominant choices in Game 1 is exactly the same as in Game 3 but Payoff level in Game 1 is more than twice as large as in Game 3. Result 2: There is no significant difference in players’ choices between Game 2 and Game 3. 7. Experiment 32 The experimental results support our second hypothesis because no difference between Game 2 and Game 3 was detected at the 5% significance level. The p-value of the Chi-square test is 0.3. The results from Game 2 and Game 3 are summarized in Table 7.2. Even though there is no significant difference, we can see higher number of risk dominant choices A in Game 2 than in Game 3. This point is in line with Relative riskiness predictions because decreasing the Relative riskiness parameter RR should result in a higher number of risk dominant choices. Moreover, considering Result 1, it seems that the Payoff dominance parameter P did not play any role in Battalio et al. (2001) and Dubois et al. (2011)’s results. Thus, their results should be solely explainable by Relative riskiness. We should, therefore, observe significant differences in our results if Relative riskiness truly matters. There are two possible explanations of non-significant differences. The first explanation might be that differences in Relative riskiness parameter RR were not large enough to induce players to change their choices. We increased the parameter RR from 41 in Game 2 to 12 in Game 3. For comparison, Dubois et al. (2011) increased the parameter RR from 14 in their Game 2 to 23 in their Game 1. The relatively small number of observations might be seen as another explanation. In total, we have 68 independent observations per game. The number of our observations is, however, higher than in experiments of Schmidt et al. (2003) (40 observations per game), Battalio et al. (2001) (64 observations per game) or Dubois et al. (2011) (64 observations per game). It may suggest that if there were significant effect of Relative riskiness, it would have been observed in our experiment. 7.3 Further Data Analyses In the following paragraphs, we focus on further data analyses of our experimental results. First, the results are regressed on participants’ characteristics, such as age or obtained degree, in order to determine whether these characteristics can explain the experimental outcome. Different types of participants are introduced together with their counts in sessions. Differences between sessions are, further, investigated. Coordination and miscoordination of participants in different games are examined as well. Next, the types of orders are discussed. In the end, the gender differences are explored. 7. Experiment 33 We regressed our results from Game 1, Game 2 and Game 3 on participants’ gender, age, obtained degree and number of visited experiments. None of our independent variables is significant at 5% significance level. The Pseudo R2 ranges from 1.1% in Game 2 to 5.08% in Game 3. We, therefore, conclude that players’ gender, age, obtained degree and number of visited experiments do not explain our experimental results well. The output from Stata is enclosed in Appendix C. For purposes of our thesis, we define four different types of participants. Participants who played payoff dominant strategies in all three rounds are called Payoff Types (P T ). Participans who played once risk dominant strategy are called rather Payoff Types (rP T ). As Safe Types (ST ) are labeled participants choosing only risk dominant strategies. Rather Safe Types (rST ) played once the payoff dominant strategy. The number and percentage of participants in particular sessions are summarized in Table 7.3. There are no significant differences in participants’ choices in single games between all three sessions according to the Fisher’s exact test. The lowest pvalue of 0.358 was obtained as difference between first and second session in Game 1. Table 7.3: Participants’ Types Session 1st Number Percentage 2nd Number Percentage 3rd Number Percentage Total Number Percentage Participiant’s Type PT 9 45% 10 41.7% 7 29.2% 26 38.2% rPT 3 15% 3 12.5% 7 29.2% 13 19.1% rST 4 20% 4 16.7% 5 20.8% 13 19.1% ST 4 20% 7 29.2% 5 20.8% 16 23.5% Participants might either coordinate or miscoordinate (mC) in their choices. There were two types of coordinations: coordination on a payoff dominant Nash equilibrium (pC) or coordination on an inferior Nash equilibrium (iC). The highest number of miscoordinations happened in Game 1. On the contrary, the lowest number of miscordinations as well as the highest number of payoff dominant coordinations happened in Game 3. In total, participants were 7. Experiment 34 more successful in coordinations in comparison with miscordinations. Table 7.4 summarizes numbers of coordinations and miscoordinations in given games. The last type of order (G3 → G2 → G1) is significantly different from third type of order (G2 → G1 → G3) at 10% significance level and from all other orders at 5% significance level. Moreover, the third type of order (G2 → G1 → G3) is significantly different from the second type of order (G1 → G3 → G2) at 10% significance level. Third and sixth type of order have relatively more payoff dominant choices than other types of orders, and their common feature is that Game 1 follows after Game 2. Therefore, we compared experimental results of Game 1 following after Game 2 with results of Game 1 following after Game 3 and with results of Game 1 as the first game in an order. We did not find any significant differences in Game 1 under these three settings. Similarly to the Game 1, Game 2 and Game 3 were compared in three different settings as well. No significant differences were found. Table 7.4: Coordination & Miscoordination Game 1 Game 2 Game 3 Total Type of coordination mC iC pC 19 4 11 15 9 10 12 7 15 46 20 36 Even though there are no significant differences in gender decision (the p-value of Fisher’s exact test is 0.3168), the percentage of woman’s payoff dominant choices is almost 10% higher than by men. Moreover, 50% of women are Payoff types in comparison with 34.6% of men. The fact that women chose payoff dominant choices in about 32 of games supports the Neumann & Vogt (2009)’s statement about risk aversion. As mentioned previously, Neumann & Vogt (2009) claim that risk aversion of participants does not predetermine them to choose risk dominant choices. Women are supposed to be more risk averse in gain domain than men (Eckel & Grossman (2008)) but women were more likely to choose payoff dominant strategies in our experiment than men. Therefore, our experimental evidence supports Neumann & Vogt (2009)’s statement. Chapter 8 Conclusion To investigate whether large changes in Payoff dominance parameter matter, two Stag-hunt games were proposed. The Payoff dominance parameter was greatly increased in Game 2 in comparison with Game 1 while all other parameters were kept equal. Moreover, another Stag-hunt game was proposed to examine the predictive power of Relative riskiness. Game 3 has the same values of all parameters except for Relative riskiness parameter as Game 2. The Relative riskiness parameter in Game 3 is twice as large as in Game 2. The results from our experiment do not confirm our first hypothesis that there will be significantly more payoff dominant choices in Game 2 than in Game 1. In fact, the opposite is true. Participants chose insignificantly more payoff dominant strategies in Game 1 than in Game 2. We attribute this fact to the random noise in games. Moreover, two other possible explanations were discussed. Considering our results and Schmidt et al. (2003)’s findings, we came to the conclusion that changes in Payoff dominance parameter P do not have any significant influence on participants’ decisions. Therefore, the results from Battalio et al. (2001) and Dubois et al. (2011) cannot be explained by Payoff dominance parameter. Our second hypothesis that there will be no significant differences in participants’ choices between Game 2 and Game 3 was confirmed by our experimental evidence. There is, however, a higher number of payoff dominant choices in Game 3 in comparison with Game 2. This fact is in line with Relative riskiness predictions. Moreover, Battalio et al. (2001) and Dubois et al. (2011)’s results are explainable with Relative riskiness when the influence of Payoff dominance was rejected. Significant differences should have been observed because of 8. Conclusion 36 the predictive power of Relative riskiness. Two potential explanation for the insignificant results are provided. Our recommendation for further research is the development of a generalized form of Relative riskiness criterion. The current form of Relative riskiness requires a strong assumption that the payoffs from risk dominant choices are not equal. This assumption decreases the applicability of Relative riskiness. For example, in three out of four Schmidt et al. (2003)’s games the Relative riskiness cannot be applied. We, therefore, suggest the development of the generalized form applicable to all types of Stag-hunt games. Bibliography Anderlini, L. (1999): “Communication, Computability, and Common Interest Games.” Games and Economic Behavior 27(1): pp. 1–37. Battalio, R., L. Samuelson, & J. V. Huyck (2001): “Optimization incentives and coordination failure in laboratory stag hunt games.” Econometrica 69(3): pp. 749–764. Bellhouse, D. (2007): “The Problem of Waldegrave.” Journal Électronique d’Histoire des Probabilités et de la Statistique [electronic only] 3(2): pp. 1–12. Bruns, B. (2010): “Navigating the Topology of 2x2 Games: An Introductory Note on Payoff Families, Normalization, and Natural Order.” CoRR 1010.4727. Carlsson, H. & E. V. Damme (1993): “Global Games and Equilibrium Selection.” Econometrica 61: pp. 989–1018. Carmichael, F. (2005): A guide to game theory. New York: Financial Times Prentice Hall, 1 edition. Casti, J. (1997): Five golden rules: great theories of 20th-century mathematics and why they matter. New York: Wiley, 1 edition. Cooper, R., D. D. Jong, R. Forsythe, & T. Ross (1992): “Forward induction in coordination games.” Economics Letters 40(2): pp. 167–172. Dubois, D., M. Willinger, & P. Van Nguyen (2011): “Optimization incentive and relative riskiness in experimental stag-hunt games.” International Journal of Game Theory 41(2): pp. 369–380. Eckel, C. C. & P. J. Grossman (2008): “Forecasting risk attitudes: An experimental study using actual and forecast gamble choices.” Journal of Economic Behavior & Organization 68(1): pp. 1–17. Bibliography 38 Feltovich, N., A. Iwasaki, & S. H. Oda (2012): “Payoff levels, loss avoidance, and equilibrium selection in games with multiple equilibria: an experimental study.” Economic Inquiry 50(4): pp. 932–952. Fischbacher, U. (2007): “z-Tree: Zurich toolbox for ready-made economic experiments.” Experimental Economics 10(2): pp. 171–178. Gallice, A. (2006): “Predicting one Shot Play in 2x2 Games Using Beliefs Based on Minimax Regret.” Fondazione Eni Enrico Mattei Working Paper 2006.31. Gallice, A. (2007): “Some equivalence results between mixed strategy Nash equilibria and minimax regret in 2x2 games.” Economics Bulletin 3(28): pp. 1–8. Gibbons, R. (1992): A Primer in Game Theory. Financial Times Prentice Hall, 1 edition. Greiner, B. (2004): “An online recruitment system for economic experiments.” Forschung und wissenschaftliches Rechnen, GWDG Bericht 63 pp. 79–93. Guyer, M. & A. Rapoport (1972): “2 x 2 games played once.” Journal of Conflict Resolution 16(3): pp. 409–431. Harsanyi, J. C. (1995): “A new theory of equilibrium selection for games with complete information.” Games and Economic Behavior 8(1): pp. 91–122. Harsanyi, J. C. & R. Selten (1988): A General Theory of Equilibrium Selection in Games. Cambridge: MA: The MIT Press, 1 edition. Huyck, J. V., R. Battalio, & R. Beil (1990): “Tacit coordination games, strategic uncertainty, and coordination failure.” The American Economic Review 80(1): pp. 234–248. Keser, C. & B. Vogt (2000): “Why do experimental subjects choose an equilibrium which is neither payoff nor risk dominant?” CIRANO Working Papers 2000s-34. Nash, J. (1950): “Equilibrium points in n-person games.” Proceedings of the national academy of sciences 36(1): pp. 48–49. Bibliography 39 Neumann, T. & B. Vogt (2009): “Do Players’ Beliefs Or Risk Attitudes Determine the Equilibrium Selections in 2x2 Coordination Games?” Working Paper University of Magdeburg 09024. Poundstone, W. (1992): Prisoner’s dilemma. New York: Anchor Books, 1 edition. Rasmusen, E. (2006): Games and Information: An Introduction to Game Theory. Malden: Blackwell Publishing, 4 edition. Romp, G. (1997): Game theory: introduction and applications. New York: Oxford University Press, 4 edition. Rydval, O. & A. Ortmann (2005): “Loss avoidance as selection principle: evidence from simple stag-hunt games.” Economics Letters 88(1): pp. 101– 107. Schelling, T. (1960): The Strategy of Conflict. Cambridge: Harvard University, 4 edition. Schmidt, D., R. Shupp, J. M. Walker, & E. Ostrom (2003): “Playing safe in coordination games:.” Games and Economic Behavior 42(2): pp. 281–299. Selten, R. (1995): “An Axiomatic Theory of a Risk Dominance Measure for Bipolar Games with Linear Incentives.” Games and Economic Behavior 8(1): pp. 213–263. Skyrms, B. (2001): “The Stag Hunt.” Proceedings and Addresses of the American Philosophical Association 75(2): pp. 31–41. Straub, P. G. (1995): “Risk dominance and coordination failures in static games.” The Quarterly Review of Economics and Finance 35(4): pp. 339– 363. Tucker, A. & R. Luce (1959): “On the Theory of Games of Strategy.” Contributions to the Theory of Games 4: pp. 13 – 42. Volek, J. (2010): Modelovánı́ a rešenı́ rozhodovacı́ch situacı́. Pardubice: Univerzita Pardubice, 1 edition. Appendix A Instructions The original instruction in Czech (Figure A.1) are on the next page followed by into English translated instruction (Figure A.2). A. Instructions II Figure A.1: Instruction Instrukce Tato část experimentu se skládá ze tří na sobě nezávislých kol. V každém kole budete náhodně přiřazeni k některému z ostatních účastníků experimentu, přičemž Vaše výplata bude záviset na rozhodnutí tohoto účastníka. V žádné z her nebude hrát s účastníkem, s kterým už jste hráli. Na obrázku níže vidíte obrazovku, která se Vám zobrazí během experimentu. Prostředí hry se skládá ze tří polí: levé pole (Vaše výplata), prostřední pole (Vaše volba) a pravé pole (Výplata druhého účastníka). Písmena W, X, Y a Z označují Vaši možnou výplatu z daného kola v závislosti na rozhodnutí druhého účastníka a w, x, y a z označují možnou výplatu druhého účastníka v závislosti na Vašem rozhodnutí. Tato písmena budou během experimentu nahrazena různými číselnými hodnotami, přičemž W=w, X=x, Y=y a Z=z. Vaše volba (A či B) označuje řádek tabulky, kde se v závislosti na volbě druhého účastníka určí Vaše výplata. Volba druhého účastníka (A či B) označuje naopak sloupec, kde v závislosti na Vašem rozhodnutí bude určena jeho výplata. Příklad Pokud v prostředním poli zvolíte volbu A a druhý hráč zvolí volbu B, tak Vaše odměna bude X (z levého pole) a odměna druhého hráče bude y (z pravého pole). Pokud v prostředním poli zvolíte volbu B a druhý hráč zvolí volbu A, tak Vaše odměna bude Y (z levého pole) a odměna druhého hráče bude x (z pravého pole). Na konci experimentu Vám budou zobrazeny Vaše výplaty z těchto tří her. Vyplacena Vám bude jedna z náhodně vylosovaných her. A. Instructions III Figure A.2: Translated instruction Instruction This part of experiment consists of three independent rounds. You will be randomly matched with one of other participants in each round. Your payoff will be based on your decision as well as on decision of other participant you will be matched with. You will not be matched twice with the same participant in any round. You can see an experimental screen in the picture below. Similar screen will be displayed to you during the experiment. The game setting is composed of three fields: left field (Your payoff), middle field (Your choices) and right field (Payoff of other participant). Letters W, X, Y and Z represents your possible payoffs from a given period depending on your decision and decision of other participant. Letters w, x, y and z represents the possible payoff of other participant based on his or her decision and your decision. It holds that W=w, X=x, Y=y a Z=z. The letters will be replaced with numerical values during the experiment. Your choice (A or B) represents the row from which you will receive your payoff depending on the other participant choice. The other participant choice (A or B) represents the column from which you will receive your payoff depending on your decision. Example If you select choice A in the middle field and the other participant selects for choice B, then your payoff will be X (from the left field) and the other participant payoff will be y (from the right field). If you select for choice B in the middle field and the other participant select for choice A, then your payoff will be Y (from the left field) and the other participant payoff will be x (from the right field). All your payoffs from three rounds will be displayed to you at the end of the experiment. One of randomly selected periods will be paid you off. Appendix B Experimental screen The following three screenshots of Games 1, Game 2 and Game 3 represents participants’ experimental screens during the experiment. Figure B.1: Game 1 B. Experimental screen V Figure B.2: Game 2 Figure B.3: Game 3 Appendix C Output from Stata We regressed our results from Game 1, Game 2 and Game 3 on participants’ gender, age, obtained degree and number of visited experiments. G1-mfx, G2-mfx and G3-mfx represent marginal effects of particular variables. Figure C.1: Regression
© Copyright 2026 Paperzz