Inference under Stability of Risk Preferences


Speaker:Joshua Teitelbaum, Georgetown University
Date: Wednesday 26 June 2013
Time: 16.00
Location: Bateman Lecture Theatre, Building One

Further details

Economists strive to develop models that can explain behavior across manifold domains. At a minimum, we ask that a model's explanatory power extend across contexts that are essentially similar. Stated more formally, we require that a model satisfy a criterion of stability: a single parameterization of the model should be consistent with observed behavior in closely related contexts. One can treat stability as a testable hypothesis. In the case of models of consumer choice, we can combine the stability criterion with revealed preference arguments to test the hypothesis that consumers' preferences - as represented by the model and revealed by their choices - are stable across similar decision contexts. Barseghyan et al. (2011) and Einav et al. (2012) take this approach to examine the stability of risk preferences, investigating whether there exists a single parameterization of the expected utility model that is consistent with data on insurance choices across multiple lines of coverage. Both found that most consumers (more than two thirds) do not exhibit stable risk preferences under the expected utility model. Yet Einav et al. (2012) also found that consumers' choices are rank correlated across contexts, suggesting that their risk preferences have a domain-general component but are not well represented by the expected utility model.

Motivated in part by this work, Barseghyan et al. (2012) [hereafter, BMOT] used data on households' insurance choices to estimate a generalization of the expected utility model that allows for "generic" probability distortions. The probability distortions in the BMOT model are generic in the sense that they can arise from a number of sources, including systematic risk misperceptions, rank-dependent probability weighting (Quiggin 1982), K½oszegi-Rabin loss aversion (K½oszegi and Rabin 2006, 2007), and Gul disappointment aversion (Gul 1991); see BMOT for a discussion. Based on their estimates, BMOT concluded that probability distortions in the form of substantial overweighting of claim probabilities play an important role in explaining the data. We take a different approach in this paper. Rather than treat stability as a testable hypothesis, we exploit the stability criterion to conduct inference on the structure of households' risk preferences, as represented by the probability distortion model and revealed by deductible choices in three lines of property insurance. We take a partial identification approach (Manski 2003), making
minimal additional assumptions and adding these assumptions sequentially in order to transparently show the role that each plays in sharpening the inference. Under this approach, we need make no assumptions about the relationship between observed heterogeneity and risk preferences, nor about the distribution of unobserved heterogeneity in risk preferences. The idea is simply to use revealed preferences arguments to bound the model parameters and to exploit the stability criterion and other minimal assumptions - all but one of which amount to shape restrictions on the utility and probability distortions functions- to sharpen the inference. The more choices we observe per household, the more precise the inference we can make about the household's risk preferences.