Posterior probability of the Likelihood Ratio and (Fractional) Bayes Factor: new theoretical relations and practical uses

TL;DR

This paper establishes new theoretical relations between the posterior probability of the likelihood ratio and Bayes factors, and demonstrates their practical application in exoplanet detection using Markov Chain methods.

Contribution

It introduces novel theoretical relations linking the likelihood ratio's posterior probability with Bayes factors and proposes optimized detection procedures based on ROC curves.

Findings

Posterior mean of LR equals BF in simple vs composite tests.

Posterior standard deviation of LR indicates detection significance.

Optimized detection procedures improve exoplanet detection performance.

Abstract

In the simple vs composite hypothesis test with a proper prior, the Bayes Factor (BF) is shown to be the posterior mean of the Likelihood Ratio (LR). Therefore, the posterior standard deviation of the LR or rather its posterior cumulative density function can be used to indicate the significativity of a detection by the BF and this detection procedure can be computed from a single Markov Chain. It is applied and compared for exoplanet detection. The previous statistics can be expressed from the Fractional BF (FBF) \cite{ohagan95} and the Probability distribution of the LR (PLR) \cite{aitkin97}. Two properties of the PLR related to the GLRT are noted and a procedure to optimize the PLR and the FBF two-parameters detectors according to their ROC curves is proposed. The performances of all tests are compared.