Agent-based modeling and experimental economics: exploring the behaviors behind the outcomes

A series of experiments was conducted on a single set of subjects to investigate learning and correlated responses. Altruism was inferred from an ultimatum game, learning was inferred by comparing the results of two successive beauty contest experiments, and iterative thinking was inferred from the beauty contest results. This paper presents a process for numerically modeling the experimental process, then developing an agentbased model (ABM) with a parameter space informed by the numerical analysis. A single candidate ABM is presented for altruism, and another for iterative thinking. Two simple behavior rules were able to reproduce the iterative thinking and altruism results, nearly identically to the experimental results in the case of altruism. Background In November 2004, a series of experiments was conducted at University of New Mexico to examine any link between iterative thinking and altruism. Iterative thinking was captured in a beauty contest, while altruism was captured in an ultimatum game. Additionally, the beauty contest was run twice, with a brief illustrative example between them, and the results correlated by subject as a measure of learning. There were four parts to the experiment: 1. Beauty contest A 2. Example beauty contest 3. Beauty contest B 4. Ultimatum game The beauty contests were intended to reveal iterator clusters – groups of subjects with similar depths of iterative thinking. The instructions were to guess two-thirds of the group mean. A zero-order iterator is one who picks the anticipated mean (the experiment revealed two consensus values for the anticipated mean). A first-order iterator picks two thirds of the anticipated mean. An order-two iterator picks two-thirds of that value, and so on. 1 The experiment was designed and conducted by Curt Shepherd while at the University of New Mexico Department of Economics. Results used by permission. Analysis of Experiment Data The analysis that follows might typically be relegated to an appendix or omitted altogether. In computer modeling, however, defining the problem well is typically the biggest hurdle. The first step, therefore, is thorough analysis of the data. The goal is not rigorous analysis of the experimental result, but rather an exploration of evidence of causal behaviors that could explain the results. First, the beauty contest data are examined. The initial approach is to assess the combined data from both beauty contests combined, assuming that this is similar to two different data sets. Figure 1 is a histogram of the raw data.