Rationality and Error in Individual Choice Data: A Revealed Preferences Approach (Job Talk-8:45am)

Abstract: The assumption of rationality is used to infer preferences from observed choices, but classic economic theory provides scant guidance when choice data has error. Methods to estimate preferences from noisy data inevitably invoke additional assumptions on those very preferences. This paper presents a procedure to detect and measure error in an individual’s observed choice data when the individual has an underlying rational choice process that has been contaminated with random implementation errors. Using a single individual’s choices over many menus, I construct an observed revealed preference relation, and prove it is a random graph whose acyclicity is equivalent to rationality. Exploiting the structure in the graph produced by the contaminating errors, I devise a classifier able to detect which observations are errors and an estimator to measure the rate at which errors occur. These two methods can be applied to any dataset in which an individual makes constrained choices from a sequence of overlapping non-identical menus, regardless of the choice environment. I apply the method to a benchmark dataset of choices observed in the lab (Choi et al. 2007) and show that most individuals have error rates between 5% and 14.5% (interquartile range). I show that three existing measures of goodness-of-fit for rationality, which are often used as proxies for error estimates, are not robust, not identified, or biased when choices are observed with error