Economic Analysis of Data from Laboratory Experiments: Endogeniety, Attrition, Subject Heterogeniety, and Interdependencies

本建议解决了一些涉及经济学实验室实验数据统计分析的问题。经济学中的实验工作对该领域越来越重要，但与现有的机会和提出的研究问题相比，对这些实验数据的统计分析仍然相对原始。我们调查的目的既有方法论的，也有实质性的：方法论的意义在于展示现代统计（计量经济学）技术在数据分析问题上的应用，这些问题是实验者关注的中心问题，并将这些技术应用于实验室实验数据提出的特殊问题（和机会）。实质性的是，一些特定的问题已经在文献中得到关注，所以这里提出的新实验，以及相应的数据分析，应该为手头的问题提供新的见解。就若干问题提出了工作建议。(1)在普通价值（或矿物权）拍卖实验中，存在显著的自我选择效应，因为落入赢家诅咒的竞标者会损失金钱，往往会破产，并以很少的钱结束实验。因此，他们倾向于自我选择不参加后续的实验环节，这可能导致对更有经验的竞标者的行为得出的结论存在重大偏差。计划是测量偏差，并探索不同的方法来处理它，从统计和实验设计的角度。(2)在经济实验中，一种常见的方法是在给定的实验阶段内提供有回报的重复试验，以使受试者获得有关经济偶然性的经验。这导致受试者从实验中积累的收益随着时间的推移而被忽略，或者当研究未能解释这些收益变化的内生性时。这反过来又会导致拍卖结果分析中的重要偏差。我们计划确定为什么当经济理论表明代理人不会受到这种影响时，他们会受到这种影响，并扩展开发的实验方法，以研究采矿权拍卖实验中收益对现金余额影响的影响。(3)长期以来，经济学家一直怀疑，经济实验中的表现至少部分反映了实验对象在实验室之外的经历以及他们的人口特征。我们将通过收集相关的人口统计和经验数据来系统地研究这一问题，并借鉴在实地数据统计分析中发展起来的方法。(4)在许多经济学实验中，个体受试者的观察结果与一组受试者在给定实验期间的重复观察结果之间存在潜在的相互依赖性。我们计划在许多不同的环境中探讨这个问题，并设计处理这个问题的方法。(5)在对经济实验中的行为进行统计评估时，研究人员通常假设所有或大多数受试者的反应系数是恒定的。在许多情况下，这可能是不正确的。我们计划研究在这种情况下个体间这种反应异质性的解释方法。

This proposal addresses a number of issues involved in the statistical analysis of data from laboratory experiments in economics. Experimental work in economics is of growing importance to the field, but the statistical analysis of data from these experiments is still relatively primitive compared to the opportunities available and the research issues posed. The purpose of our investigations is both methodological and substantive: Methodological in the sense of demonstrating applications of modern statistical (econometric) techniques to data analysis issues of central concern to experimenters, and refining applications of these techniques to the special questions (and opportunities) that laboratory experimental data pose. Substantive in that a number of the particular issues addressed have been of concern in the literature, so that the new experiments proposed here, and the corresponding data analysis, should provide new insights into the issues at hand. Work is proposed on a number of issues. (1) In common value (or mineral rights) auction experiments there is a significant self-selection effect as bidders who fall prey to the winner's curse lose money, often go bankrupt, and end the experiment with very little money. As a consequence they tend to self-select out of returning for subsequent experimental sessions which can lead to important biases in the conclusions reached regarding the behavior of more experienced bidders. The plan is to both measure the bias and to explore different ways of dealing with it, both statistically and in terms of experimental design. (2) One common method in economic experiments is to provide repeated trials with payoffs within a given experimental session to provide subjects with experience with the economic contingencies involved. This leads to subjects accumulating earnings from the experiment over time whose effect is typically ignored, or when studied fails to account for the endogeneity of changes in these earnings. This in turn can lead to important biases in the analysis of auction outcomes. We plan to determine why agents are subject to such effects when economic theory suggests that they will not be, and to extend the experimental methodology developed to study the impact of earnings on cash balance effects in mineral rights auction experiments. (3) Economists have long suspected that performance in economic experiments reflects, at least in part, the experiences of subjects outside the lab as well as their demographic characteristics. We will study this systematically by collecting relevant demographic and experiential data on subjects borrowing methods developed in the statistical analysis of field data. (4) In many economic experiments there are potential interdependencies between individual subject observations associated with repeated observations of a group of subjects within a given experimental session. We plan to explore this issue in a number of different settings and to devise ways for dealing with it. (5) Often when statistically evaluating behavior in economic experiments investigators assume that all, or most, response coefficients are constant across subjects. This is likely to be incorrect in a number of cases. We plan to study ways of accounting for this response heterogeneity across individuals in such cases.