Robust statistical inference based on the \(C\)-divergence family (Q2330537)

The authors study a family of divergences between densities, the so called \(C\)-divergence, which includes the (generalized) power divergence family and the S-divergence family as special cases. They give results on the robustness and the asymptotic distribution, so that they can construct robust tests. To this end, they derive the first and higher order influence function, it turns out that (under some choice of parameters) the influence function is bounded and redescending, indicating the robustness of this divergence. This is also confirmed by a breakdown point reaching up to 50\% and it is illustrated by a Poisson mixture model. The asymptotic normality is shown for both discrete and continuous models. The article also studies the behavior of the \(C\)-divergence in practice, giving empirical levels and powers of a test based on the generalized power divergence in a Poisson model. The test is applied to several classical data sets like Newcomb's speed of light data.