Beyond the Domain of Direct Observation: How to Specify a Probability Distribution that Represents the “State of Knowledge” About Uncertain Inputs