Error cascades in observational learning: An experiment on the Chinos game