We ran three-round sequential bargaining experiments in which the perfect equilibrium offer was $1.25 and an equal split was $2.50. Subjects offered $2.11 to other subjects, $1.84 to "robot" players (who are known to play subgame perfectly), and $1.22 to robots after instruction in backward induction. Measures of information search showed that subjects did not look at the amounts being divided in different rounds in the correct order, and for the length of time, necessary for backward induction, unless they were specifically instructed. The results suggest that most of the departure from perfect equilibrium is due to limited computation and some is due to fairness.