We analyze what can be inferred about a game's information structure solely from the probability distributions on action profiles generated during play; i.e., without reference to special behavioral assumptions or equilibrium concepts. Our analysis focuses on deriving payoff-independent conditions that must be met for one game form to be empirically distinguished from another. We define empirical equivalence and independence equivalence. The first describes when two game forms can never be distinguished based solely on the empirical distribution of player actions. As this turns out to be difficult to characterize, we introduce the latter, which describes two game forms that imply the same minimal sets of conditional independencies in every one of their empirical distributions. Our main contribution is to identify, for an arbitrary game form, the minimal set of conditional independencies that must arise in every one of its empirical distributions. We also introduce a new graphical device, the influence opportunity diagram of a game form which facilitates verifying independence equivalence, and hence provides a simple necessary condition for empirical equivalence.
- Bayesian network,
- extensive game form,
- emprical inference
Available at: http://works.bepress.com/michael_ryall/7/