On Rationalization of Nash and Subgame-Perfect Behaviour from Observed Outcomes ∗