Modeling expert judgements for Bayesian updating (Q1086924)

Consider a Bayesian decision maker who updates his probability p of an event A, in the light of a number of expert opinions expressed as probabilities \(q_ 1,...,q_ n\) for occurrence of A, by the formula \(p^*(q)=p \Pr (Q=q| A)/\Pr (Q=q)\) assuming that the n expert opinions form a random vector \(Q=(Q_ 1,...,Q_ n)\). Suppose the decision maker is unable to specify \(\Pr (Q=q| A)\) and \(\Pr (Q=q| \bar A)\) completely, and yet seeks a formula \(p^*(q)\) satisfying the consistency condition \(p^*(q)=\Pr (A| Q=q).\) It is shown that if the decision maker can specify the mean vector \(\mu =(\mu_ 1,...,\mu_ n)\) of Q then a function \(p^*(q)\) satisfies the consistency condition if and only if it is of the form \[ \lambda_ 0p+\sum^{n}_{1}\lambda_ i(q_ i-\mu_ i) \] for some \(\lambda_ 0,\lambda_ 1,...,\lambda_ n\) satisfying certain inequalities. Also, a product formula for \(p^*(q)\) is derived under some regularity conditions for the case where the expert opinions are conditionally independent given A.