Identification and Inference of Strategic and Social Interaction Models

对不通过价格调节的相互作用的理解在许多经济现象中都很重要。例如，一家公司进入一个市场的决定可能会受到其他公司的决定的影响。在课堂上，同伴效应可能会影响结果和决定，而这些结果和决定可能会在以后的生活中发挥重要作用。还有很多其他的例子。由于各行为者之间决策和结果的相互作用，对这些现象的经验分析与对个人结果和决定的经验分析有关，但带来了超出个人决策分析的一般挑战。本研究项目探讨互动模型实证分析的方法论和实质应用。其中一个主要挑战是，一个给定的模型可能受制于一组以上的预测结果或行为(即“均衡”)。在这种情况下，不清楚如何从观测数据中检索重要的模型组件。在第一组研究中，研究人员扩展了之前的工作，即研究人员和合著者研究互动模型的识别，其中某些信息对参与者是私有的(De Paula和Tang，2011)。在这篇文献中的典型假设下，他们表明，当可能有多种解决方案并且不同的数据点对应于潜在的不同解决方案时，即使使用定义非常宽松的模型，也可以推断某人是否受到激励或劝阻(当其他人采取行动时)。作为副产品，调查人员还能够测试数据中是否存在多种解决方案。这一点很重要，因为对于目前可用的许多估计技术来说，多重性是有问题的。研究人员将这一分析扩展到几个方向。首先，尽管在标准假设下，关于检测多个解的基本见解仍然适用于动态博弈，但由于一个人目前的行为影响到未来的环境，因此交互效应提取的结果会受到影响。第二，即使在最初调查的静态环境中，连续分布的控制变量，如价格或收入，在估计时也会带来实际的复杂性。因为这些变量的不同值可能在基础经济中导致不同数量的解，所以交互作用效应的典型非参数估计聚集在协变量上，这污染了人们对解的唯一性的推断，从而污染了交互作用的符号。最后，尽管在许多应用中，人们可以很容易地给相关参与者(例如，丈夫和妻子、沃尔玛和当地商店)贴上标签，但在其他情况下，标签并不自然地出现。研究人员还调查了这如何影响互动效应的推断，并将这些方法应用于分析大学室友之间的同伴效应，这是近年来被广泛研究的一个话题。在一个相关的项目中，研究人员专注于人们选择与谁交往的经验模型(即“网络形成游戏”)。最近，经济学和社会科学中的网络现象引起了人们的广泛关注。例如，在两种环境中找到多个均衡是很常见的。由于网络形成模型与经验产业组织文献中通常研究的模型有许多相似之处，因此类似于先前在该文献分析中使用的思想和技术可以适用于社会网络的研究。为了演示在这项研究中开发的方法，使用AddHealth数据集提供了一个插图。这项研究中提出的技术将为随后的结果分析提供重要的第一步，这些结果受到个人如何相互联系的影响。例如，教育成果和信息传递。

The understanding of interactions not mediated via prices is important in many economic phenomena. For example, a firm's decision to enter a market may be affected by other firms' decisions. In the classroom, peer effects may affect outcomes and decisions which may in turn play an important role later in life. There are many other examples. Because of the interplay of decisions and outcomes across actors, the empirical analysis of these phenomena is related to the empirical analysis of individual outcomes and decisions but poses challenges beyond those typically found in the analysis of individual decision making. This research project investigates methodological aspects and substantive applications in the empirical analysis of interaction models.One of the main challenges arises as a given model may be amenable to more than one set of predicted outcomes or behaviors (i.e. "equilibria"). When this is the case it is not clear how to retrieve important model components from observed data. In a first set of studies, the researcher extends previous work where the investigator and a co-author study the identification of interaction models where certain information is private to participants (de Paula and Tang, 2011). Under typical assumptions in this literature they showed that when multiple solutions are possible and different data points correspond to potentially different solutions, whether someone is incentivized or dissuaded from taking an action (when others do) can be inferred even with a very loosely defined model. As a byproduct, the investigators are also able to test for the existence of multiple solutions in the data. This is important since multiplicity is problematic for many estimation techniques currently available. The researcher extends that analysis in several directions.First, whereas the basic insight on the detection of more than one solution remains applicable in dynamic games under standard assumptions, the results on the retrieval of interaction effects is affected since what a person does at present affects the environment in the future.Second, even in the static environment originally investigated, continuously distributed control variables such as prices or income pose practical complications in estimation. Because different values for those variables may induce a different number of solutions in the underlying economy, typical nonparametric estimates of the interaction effect aggregate over covariates that contaminate one's inference on the uniqueness of a solution and, consequently, on the sign of interaction effects.Finally, whereas in many applications one can easily label the relevant players (e.g., husband and wife, Wal-Mart and local stores), in other contexts the labels do not naturally present themselves. The researcher also investigates how this affects the inference of interaction effects and applies the methods to the analysis of peer effects among college roommates, a topic that has been studied extensively in recent years.In a related project, the researcher focuses on empirical models where people choose who to associate with (i.e. "network formation games"). Recently, a lot of interest has developed around network phenomena in Economics and the Social Sciences more generally. It is common, for example, to find multiple equilibria in both environments. Because a network formation model bears many similarities to models usually studied in the empirical Industrial Organization literature, ideas and techniques akin to those previously used in the analysis of that literature can be adapted to the study of social networks. To demonstrate the methodology developed in this research, an illustration using the AddHealth dataset is presented. The techniques proposed in this research will provide an important first step in the subsequent analysis of outcomes which are influenced by how individuals connect to each other. Examples include educational outcomes and information transmission.