Two partners in a venture can each withdraw at successive dates. The venture grows as long as both stay, and whoever withdraws first takes the larger share of what has accumulated. Both would gain by staying to the end, if each could be sure that the other would not leave one step earlier.
The game tests backward induction in its purest form. It shows how an argument that is compelling at the last node loses its hold as it is carried back to the first, and it raises the question of what a player should believe after observing a move that the theory says a rational opponent would not make. The outcome also demonstrates the gains that mutual trust makes available.
In the game introduced by Rosenthal (1981), the mover at the final node does better by taking than by passing. Knowing this, the mover at the node before does better by taking there, and the argument continues to the first node, where the first player takes at once. The outcome is the unique subgame perfect equilibrium, and both players earn far less than they would by passing to the end. Aumann (1995) proves that common knowledge of rationality implies this outcome in games of perfect information.
The assumption is demanding, because a player who is passed to has seen a move that contradicts it and must decide what to make of the opponent. McKelvey and Palfrey (1992) account for observed play with a model in which a small share of players always pass, and in which the remaining players, uncertain of whom they face, find it worthwhile to pass in the early nodes and take with rising probability as the end approaches.
Few games end where the theory predicts, as the experiments of McKelvey and Palfrey (1992) established. There, with piles that doubled at every pass as they do in the default here, 7 percent of four-move games and fewer than 1 percent of six-move games ended at the first node. Most ended in the middle, and between 1 and 5 percent were passed to the end.
Expertise in backward induction does not settle the matter. Palacios-Huerta and Volij (2009) report that 69 percent of chess players, and all Grandmasters, stopped at the first node. Levitt, List and Sadoff (2011) found the opposite in a similar population, where fewer than 4 percent of games between chess players ended at the first node and no Grandmaster stopped there, although a majority of pairs solved the simpler of two other games by backward induction.