Equilibrium selection in coordination games has generated a large literature. Kandori, Mailath and Rob (1993) and Young (1993) studied dynamic models of aggregate behaviour in which agents choose best responses to observations of population play. Crucially, infrequent mistakes (`mutations`) allow agents to take actions contrary to current trends and prevent initial configurations from determining long run play. An alternative approach is offered here: Harsanyian trembles are added to agents` payoffs so that with some probability it is optimal to act against the flow of play. The long run distribution of population behaviour is characterised - modes correspond to stable Bayesian Nash equilibria. Allowing the variance of payoff trembles to vanish, via a purification process, a single equilibrium is played almost always in the long run. Kandori et al and Young show that the number of contrarian actions required to escape an equilibrium determines selection; here, the likelihood that such actions are taken is of equal importance.
