A Monte Carlo method for an objective Bayesian procedure (Q2640335): Difference between revisions

Often a statistical model is described by the likelihood function L(\(\theta\),Y) for a given set of data Y and an unknown parameter vector \(\theta\). In the case of an ill-conditioned situation the function \(\ln L(\theta,Y)-Q(\theta,\tau)\) is introduced, where Q represents a set of penalties and \(\tau\) is the vector of respective weights for the penalties. Consider the prior distribution \(\pi\) (\(\theta\) /\(\tau\)) with \(\pi (\theta /\tau)=\exp \{-Q(\theta,\tau)\}/\int \exp \{- Q(\theta,\tau)\}d\theta.\) Then \(\Lambda (\tau;Y)=\int L(\theta,Y)\pi (\theta,\tau)d\theta\) is the Bayesian likelihood of \(\tau\), and is useful to obtain the optimal hyper-parameter \(\tau\) which maximizes \(\Lambda\) or its logarithm. A Monte Carlo integration method is described and is used to determine \(\Lambda\) (\(\nabla,Y)\) and some numerical examples are discussed.