Social preferences play an important role in a large variety of domains, such as labor market relations, the provision of public goods, and redistributive politics. It is therefore important to know more about how these preferences are distributed amont the general population.

We elicit distributional preferences from a large population sample using dictator games in which participants allocate money between themselves and others. Our experiment follows a panel structure with two measurement waves three years apart, and fresh subjects entering for wave 2. Based on these data, we characterize heterogeneity of distributional preferences using a novel Bayesian nonparametric method. The method does not impose assumptions on the structure of preferences. Moreover, it explicitly allows for the full continuum of possible type distribution in the population, ranging from a single representative agent (all subjects are the same) to all agents being different.

We identify three fundamentally distinct and empirically relevant behavioral types in the data. The tractable and parsimonious characterization consists of 1. a majority type who is generally inequality averse, 2. a somewhat smaller, but still large type who exhibits an altruistic concern for those being worse off, and 3. a minority type who is predominantly selfish. The preference characterization is strikingly stable over time, and it is robust across different data sets (those who participated in both waves, those who dropped out after the first wave, and those who entered in the second wave).

In a prediction exercise, we demonstrate that the type characterization comprises relevant information about people’s behavior within and across contexts. The three types capture the essential heterogeneity in distributional preferences in that they predict much better than a representative agent; they only predict a little worse than individual-level estimates, however. We contrast type-level predictions with predictions obtained via a state-of-the-art machine learning algorithm.

Co-authors : Ernst Fehr (Zurich), Julien Senn (Zurich)

Practical information