When tension is just a fluctuation: How noisy data affect model comparison

Abstract

Summary statistics of likelihood, such as Bayesian evidence, offer a principled way of comparing models and assessing tension between, or within, the results of physical experiments. Noisy realisations of the data induce scatter in these model comparison statistics. For a realistic case of cosmological inference from large-scale structure, we show that the logarithm of the Bayes factor attains scatter of order unity, increasing significantly with stronger tension between the models under comparison. We develop an approximate procedure that quantifies the sampling distribution of the evidence at a small additional computational cost and apply it to real data to demonstrate the impact of the scatter, which acts to reduce the significance of any model discrepancies. Data compression is highlighted as a potential avenue to suppressing noise in the evidence to negligible levels, with a proof of concept demonstrated using Planck cosmic microwave background data.