MODELING PROBABILISTIC AGENT OPINION

A single agent (an individual, expert or model) provides a decision maker with probabilistic information partially or completely describing the agent's opinion about a collection of uncertain quantities or events. This paper discusses ways in which the decision maker may model the agent's opinion to provide rules for updating his own probability for a related event or random quantity of particular interest. Concepts discussed include the relevance of the agent's information and experience, the accord between the agent and decision maker in terms of common or conflicting information and calibration of probability assessments. New theory develops and extends that of Genest and Schervish, requiring only a partial specification of the decision maker's prior over the agent's opinion. Several illustrative examples are developed.