Bandit cascade: A test of observational learning in the bandit problem

I conduct an experimental investigation of observational (social) learning in a simple two-armed bandit framework where the models are based on Bayesian reasoning and non-Bayesian count heuristics providing different predictions. The agents can choose between two alternatives with different probabilities of providing a reward. They must make their choice in order to see the outcome and act in a sequence. They can base their decision on the choices of the predecessors and the outcomes of their own choice. The results of the experiment follow neither Bayesian Nash Equilibrium nor Naïve herding model (BRTNI): Subjects follow and cascade on choices that contain no information about the state of the world, and, therefore, sustain losses when learning from others. I also test the Quantal response equilibrium and the robustness of this theory.