TL;DR
This paper challenges a core assumption in game theory: that simply shifting all payoffs up or down shouldn't change behavior. Through a human-subjects experiment, they show that people do react to these shifts, particularly when they turn potential gains into losses, suggesting "loss avoidance" influences strategic choices in games with multiple equilibria.
What Problem Does This Paper Address?
Game theory often assumes that adding or subtracting a constant from all payoffs in a game should not alter players' decisions. This is based on the idea that payoffs directly reflect utility. However, when payoffs are actual monetary amounts in experiments, this equivalence isn't guaranteed. The paper investigates whether changes in payoff levels, specifically those that flip positive outcomes to negative ones (or vice versa), influence behavior. This is particularly relevant for games with multiple Nash equilibria, where standard theory doesn't predict which equilibrium will be chosen, and where payoff shifts might offer a behavioral explanation for equilibrium selection.
Methods
The researchers designed an experiment using three versions of the "Stag Hunt" game. These versions are game-theoretically identical but differ in their payoff levels, creating scenarios where payoffs can be positive, negative, or a mix. They employed four experimental treatments:
- One-shot (O): Subjects played each game once with publicly announced payoffs, allowing for deductive reasoning before experience.
- Complete Information (C): Subjects played each game 40 times, with payoffs always known, and were randomly matched with opponents. This treatment allows observation of learning and how choices evolve over time.
- Random Matching, Limited Information (R): Similar to C, but subjects did not know the payoff matrices upfront, only receiving information to infer them.
- Fixed Pairs, Limited Information (F): Similar to R, but subjects played all 40 rounds against the same opponent.
The core of their method was to track the frequency of choosing the "risky" action (which could lead to higher payoffs but also potential losses) versus the "safe" action across these different games and treatments.
Datasets
Not discussed in the paper. The study uses human subjects playing designed games.
Key Insights and Ideas
- Payoff Levels Matter: Contrary to a standard game-theoretic assumption, shifting payoff levels does influence behavior, especially when it involves potential losses.
- Loss Avoidance is Real: The experiment provides evidence for two types of loss avoidance:
- Certain-loss avoidance: Players tend to avoid strategies that guarantee a loss, even if an alternative might lead to a gain.
- Possible-loss avoidance: Players tend to avoid strategies that carry a risk of loss, preferring alternatives that offer a sure gain.
- Learning and Experience: While initial behavior might show subtle differences, loss avoidance tendencies become more pronounced and can even increase over time as subjects gain experience with the games.
- Implications for Equilibrium Selection: These behavioral biases offer a potential explanation for why certain equilibria are chosen over others in games with multiple solutions, going beyond purely rational predictions.
Key Results
- Initial Behavior: In one-shot play (O treatment), subjects showed a significant tendency towards the risky action in the low-payoff game (SHL) compared to the high-payoff game (SHH), supporting certain-loss avoidance. However, initial behavior did not show clear evidence of possible-loss avoidance between SHH and the medium-payoff game (SHM).
- Learning Over Time: In repeated play treatments (C, R, F), differences in behavior across the games became more pronounced.
- In the complete information treatment (C), loss avoidance effects were present but small, with some decline in risky action over time.
- In the limited information treatments (R and F), subjects increasingly favored the safe action in the high and medium-payoff games, while the risky action remained more prevalent in the low-payoff game. This learning was often quicker when playing against the same opponent (F treatment).
- Magnitude of Loss Avoidance: The impact of loss avoidance did not necessarily diminish with experience; in some cases, it appeared to increase over the course of the experiment.
Limitations and Future Work
The authors acknowledge that more research is needed to fully understand loss avoidance. They suggest that future work should incorporate these behavioral tendencies into models of decision-making, especially in situations involving both potential gains and losses. The paper also notes that while differences were observed, they were sometimes small, and the learning dynamics could be complex, with some effects decreasing in later rounds.
Summary updated