Hypothesis testing near singularities and boundaries (Q85619)

Typically, the asymptotic \(\chi^2\) distribution of the likelihood ratio statistic is used as an approximation in hypothesis testing. However, this approximation breaks down at (and near) boundaries and singularities of the relevant parameter spaces. In the present paper, the authors propose an alternative approximation for the distribution of the likelihood ratio statistic, and give detailed investigation of this approximation in two models in the setting of evolutionary trees. This includes investigation of the regions of the parameter space in which the standard \(\chi^2\) approximation is of a sufficiently high quality. The distribution the authors propose depends on both the sample size and the true parameter values; in light of this, they also investigate (using simulations) hypothesis testing in the presence of nuisance parameters.