Stable Randomization

We design a laboratory experiment to identify whether a preference for randomization defines a stable type across different choice environments. In games and individual decisions, subjects face twenty simultaneous repetitions of the same choice. Subjects can randomize by making different choices across the repetitions. We find that randomization does define a type that's predictable across domains. A sizable fraction of individuals randomize in all domains, even in questions that offer a stochastically-dominant option. For some mixers, dominated randomization is responsive to intervention. We explore theoretical foundations for mixing, and find that most preference-based models are unable to accommodate our results.