Experimental Studies of Incentive Effects, Expertise Effectsand Context Effects in Signalling Games

信号博弈涉及代理之间的战略交互，其中市场的一方对市场另一方的代理的质量特征具有不完全信息。信号博弈模型已被广泛应用于从营销到教育决策再到大公司之间的战略互动的各种经济环境。几乎所有有趣的信号博弈都有多个均衡(稳态)，因此必须引入某种均衡细化来缩小均衡集的范围，以便预测行为。经济理论家根据不同类型的代理人可能采取或可能不采取的看似合理的行动，得出了许多均衡修正。虽然早期的实验对信号游戏的改进提供了令人鼓舞的支持，但后来的研究发现，即使是这些标准中最弱的也有违规行为。这些后来的研究人员得出结论，受试者最好被建模为拥有天真的初始信念，只是慢慢地学习如何发挥平衡。也就是说，信号博弈改进(和一般的形式博弈论)未能产生对受试者行为的准确预测，因为对理性的潜在假设太强了。最近对我们的一项研究(Lo，Cooper，Kagel和Gu，1997)提出了几个理由，为什么不应该太快地放弃信号博弈的改进。我们发现，将所有收益增加五倍显著增加了战略博弈的频率，这可能是由于较高的收益刺激了更高的演绎推理。这表明，在其他实验中，均衡精炼的失败可能是低收益的函数。我们还比较了天真的受试者(上海中国纺织大学的本科生)和有经验的受试者(来自上海地区的纺织企业经理)的行为。这种比较是使用通用术语进行的实验(例如，公司经理是A玩家，中央计划者是B玩家，等等)。对于情景实验(旨在唤起经理们可能会遇到棘轮效应的现场环境的指令)。虽然在一般性会议上，幼稚的受试者和专家受试者之间没有发现显著的差异，但LCKG观察到，背景的存在显著提高了专家受试者的策略发挥水平，但对幼稚的受试者没有实质性的影响。LCKG假设，专家可能会学习比天真的玩家更演绎地思考，但可能只有在有一些背景作为触发因素的情况下才能利用这种专业知识。因此，信号游戏的改进在实验室实验中可能表现不佳，要么是因为受试者太天真，要么是因为受试者缺乏允许他们从其他环境中调用专业知识的背景。此外，在认知心理学文献中所报道的语境效应的存在，以及我们从认知心理学文献中了解到的语境的作用，表明了语境和学习之间的一些可能的交互作用，而到目前为止，经济学文献几乎完全忽略了这些作用。为了探索这些问题，我们的策略是构建一款新的游戏-大学招生游戏-具有美国大学生群体(作为实验对象的最方便的人群)熟悉的背景框架，并对这款游戏进行新的实验。利用这个游戏的几个变体，我们研究了环境、激励和专业知识分别和共同对学习和均衡的出现产生了什么影响。除了使用更传统的实验方法外，我们还采用了一种新的团队合作方法来收集受试者的言语数据；通过分析这些数据，我们希望获得关于受试者推理过程的直接证据，以及这些过程是如何受到实验处理的影响的。

Signaling games involve strategic interactions between agents where one side of a market has incomplete information about the quality characteristics of agents on the other side of the market. Models of signaling games have been applied to a wide variety of economic settings from marketing to educational decisions to strategic interactions between large corporations. Virtually all interesting signaling games have multiple equilibria (steady states), making it essential to introduce some sort of equilibrium refinements to narrow down the set of equilibria in order to predict behavior. Economic theorists have derived a number of equilibrium refinements based on plausible actions that different types of agents may or may not take. Although early experiments gave encouraging support for signaling-game refinements, later studies found violations for even the weakest of these criteria. These later researchers concluded that subjects could best be modeled as having naive initial beliefs, only slowly learning to play an equilibrium. That is, signaling game refinements (and formal game theory in general) failed to generate accurate predications of subjects' behavior because the underlying assumptions of rationality were too strong. A recent study of ours (Lo, Cooper, Kagel and Gu, 1997) suggests several reasons why the signaling game refinements should not be discarded too quickly. We find that scaling up all payoffs by a factor of five significantly increased the frequency of strategic play, presumably due to heightened deductive reasoning stimulated by the higher payoffs. This suggests that the failure of equilibrium refinements in other experiments may be a function of low payoffs. We also compared behavior by naive subjects (undergraduate students from China Textile University in Shanghai) with the behavior of more experienced subjects (textile firm managers from the Shanghai area). This comparison was made for experiments using generic terminology (e.g. firm managers are `A players,` central planners are `B players,` etc.) and for experiments in context (instructions designed to evoke the field settings in which managers would have encountered the ratchet effect). While no significant differences were found between naive and expert subjects in generic sessions, LCKG observed that the presence of context significantly increased the level of strategic play for expert subjects but did not substantially affect naive subjects. LCKG hypothesize that experts may learn to think more deductively than naive players, but may only be able to draw upon this expertise if there is some context to act as a trigger. Thus, signaling game refinements may do poorly in laboratory experiments either because subjects are naive or because subjects lack the context which allows them to call on expertise from other settings. Further, the existence of context effects reported in LCKG, in conjunction with what we know about the role of context from the cognitive psychology literature, suggests a number of possible interaction effects between context and learning that have so far been almost totally ignored in the economics literature. To explore these issues, our strategy is to construct a new game - the college admissions game - that has a contextual framework familiar to the US college student population (the most convenient population to use as experimental subjects) and to conduct new experiments with this game. Using several variations of this game, we study what effect context, incentives, and expertise, separately and together, have on learning and the emergence of equilibrium. In addition to using more traditional experimental methods, we also employ a new team-play methodology for gathering verbal data from subjects; by analyzing this data, we hope to gain direct evidence about the reasoning processes of subjects and how these processes are affected by the experimental treatments.