Lab Meeting 31.10.2017 - Adrian Bruhin. HEC Lausanne

Expected Utility Theory (EUT) fails as a descriptive model, as subjects’ choices under risk systematically deviate from EUT’s predictions. These systematic deviations have spurred the development of various behavioral theories of choice under risk. Prospect Theory (PT), the most prominent behavioral alternative to EUT, relies mostly on reference dependence and diminishing sensitivity to explain how the deviations from EUT come about. In contrast, Salience Theory (ST), a recently proposed contestant, is based on local thinking and salience. It postulates that, due to cognitive limitations, subjects are local thinkers and focus their attention predominantly on salient states – i.e. states in which one payoff stands out relative to the payoffs of the alternatives. PT and ST often make similar predictions. However, ST offers some advantages over PT, as it naturally extends to deterministic consumer choice, describes the counter-cyclicality of risk premia on financial markets, and explains violations of transitivity that lead to preference reversals.

In this paper, we experimentally test ST by exposing 283 subjects to a series of lottery choices that may trigger the Allais’ paradox. To reliably discriminate between PT and ST, we exploit the fact that the valuation of a lottery is context-free in PT while it is context-dependent in ST. We vary the choices’ context by altering the correlation structure of the lotteries’ payoffs. This allows us to do three things. First, we provide non-parametric and structural evidence on the two theories’ descriptive performance at the aggregate level. Second, we account for individual heterogeneity in a parsimonious way and use a finite mixture model to classify subjects into EUT-, PT-, and ST-types. Third, we perform out-of-sample predictions about the frequency of preference reversals in a set of additional choices to assess the validity of our classification of subjects into types.

Our experiment yields three main results. First, at the aggregate level, PT remains the best-fitting model while EUT and ST are both rejected. Second, however, the finite mixture model reveals considerable individual heterogeneity: PT is the best-fitting model for just 38% of subjects, while ST and EUT are the best-fitting models for 34% and 28% of subjects, respectively. Third, our out-of-sample analysis confirms this classification of subjects into types, as the ST-types exhibit significantly more preference reversals in the additional choices than the PT- and EUT-types.