Controlling risks under different loss functions: The compromise decision problem (Q5903888)

Two statisticians are presented with the problem of agreeing to use a single decision procedure from D, the collection of all such decision procedures. When each statistician is a Bayesian a ``Bayes compromise problem'' is to find a decision procedure which minimizes \(r_ i(\delta,\Pi)\) subject to \(r_ i(\delta,\Pi)\leq K_ j\) where (i,j), the roles of the statisticians, is fixed at (1,2) or (2,1) and \(K_ j\) is a fixed constant satisfying \[ r_ j(\delta_ j^ B,\Pi_ j)\leq K_ j\leq r_ j(\delta_ i^ B,\Pi_ j). \] Here \(r_ i(\delta,\Pi_ i)\) denotes the expected risk of \(\delta\) with respect to the prior distribution \(\Pi_ i\) on the parameter space for loss \(L_ i\), and \(\delta_ 1^ B\) and \(\delta_ 2^ B\) denote the Bayes procedures. The ``\(\lambda\)-Bayes compromise'' problem is to find a decision procedure which minimizes \[ \lambda r_ 1(\delta,\Pi_ 1)+(1- \lambda)r_ 2(\delta,\Pi_ 2) \] where \(\lambda\) is a fixed constant (0\(\leq \lambda \leq 1)\) ``Minimax-Bayes compromise'' problem and ``\(\lambda\)-minimax-Bayes compromise'' problem are defined similarly. The author gives a theory characterizing solutions to Bayes compromise and minimax compromise problems.