Improved Inference for the Signal Significance (1609.00752v2)

Abstract: We study the properties of several likelihood-based statistics commonly used in testing for the presence of a known signal under a mixture model with known background, but unknown signal fraction. Under the null hypothesis of no signal, all statistics follow a standard normal distribution in large samples, but substantial deviations can occur at low sample sizes. Approximations for respective $p$-values are derived to various orders of accuracy using the methodology of Edgeworth expansions. Adherence to normality is studied, and the magnitude of deviations is quantified according to resulting inflation or deflation. We find that approximations to third-order accuracy are generally sufficient to guarantee $p$-values with nominal false positive error rates in the five sigma range ($p$-value $= 2.87 \times 10^{-7}$) for the classic Wald, score, and likelihood ratio (LR) statistics at relatively low samples. Not only does LR have better adherence to normality, but it also consistently outperforms all other statistics in terms of false negative error rates. The reasons for this are shown to be connected with high-order cumulant behavior gleaned from fourth order Edgeworth expansions. Finally, a conservative procedure is suggested for making finite sample adjustments while accounting for the look elsewhere effect with the theory of random fields (a.k.a. the Gross-Vitells method).