Prior distributions for random choice structures (Q2437251): Difference between revisions

This paper is about nonparametric Bayesian analysis of random choice. Consider a master set \(T\) with \(n\) elements and call any collection of distributions on the possible subsets \(A\) of \(T\) a random choice structure. The distributions are interpreted as choice probabilities that arise in choice problem \(A\). The economic interpretation is either random choice by an individual or choices by members of a population with heterogeneous preferences. The authors propose a class of flexible, nonparametric priors over choice structures. This class implies prior probabilities of features of choice systems, notably axioms like the Marshak triangle inequality, different forms of stochastic transitivity, regularity (a stochastic property \(\alpha\)) and the event that choices are rationalizable by a distribution over preference orderings. The latter probability is small for most values of the prior's hyperparameter, illustrating that the set of rationalizable choice systems is small in the universe of choice systems. A limitation is that computations are only practical for rather small \(n\) (up to \(n=5\) in the paper). The authors leave to future research the computation of Bayes factors for axioms given data. Readers looking for Bayesian approaches to related problems might find [\textit{J. Geweke}, ``Nonparametric Bayesian modelling of monotone preferences for discrete choice experiments'', J. Econometrics 171, No. 2, 185--204 (2012; \url{doi:10.1016/j.jeconom.2012.06.003})] of interest. For frequentist approaches, consider [\textit{C. F. Manski}, ``Partial identification of counterfactual choice probabilities'', Int. Econ. Rev. 48, No. 4, 1393--1410 (2007; \url{doi:10.1111/j.1468-2354.2007.00467.x}); \textit{Y. Kitamura} and \textit{J. Stoye}, ``Nonparametric analysis of random utility models: testing'', CeMMAP working papers CWP36/13, London: Institute for Fiscal Studies (2013), \url{http://www.cemmap.ac.uk/wps/cwp361313.pdf}].