Chernoff index for Cox test of separate parametric families (Q1747731)

Let \(X_1, \ldots, X_n\) denote stochastically independent, identically distributed and observable random variables, where \(X_1\) possesses the density \(f\) with respect to some measure \(\mu\). The authors are concerned with the test problem \[ H_0: f \in \{g_\theta: \theta \in \Theta\} \text{ versus } H_1: f \in \{h_\gamma: \gamma \in \Gamma\}, \] where \(\{g_\theta: \theta \in \Theta\}\) and \(\{h_\gamma: \gamma \in \Gamma\}\) denote two separate parametric families of \(\mu\)-densities. To this end, they consider the (generalized) likelihood ratio statistic \[ \text{LR}_n = \frac{\max_{\gamma \in \Gamma} \prod_{i=1}^n h_\gamma(X_i)}{\max_{\theta \in \Theta} \prod_{i=1}^n g_\theta(X_i)}. \] They analyze the asymptotic efficiency of the resulting test by means of (a generalized version of) its Chernoff index, which is defined as the exponential decay rate of the maximal type I and type II error probabilities. In particular, the authors derive the analytic form of this quantity. As extensions, they analyze the asymptotic behavior of the corresponding Bayes factor, consider the case of more than two families, and derive results for possibly nonseparated families. The latter case is relevant for variable selection problems in (generalized) linear models. Finally, some numerical examples are provided.