A generalized divergence for statistical inference (Q2405123)

For the estimation of parameters in a parametric distribution based on an independent sample, several robust procedures were proposed by minimizing various measures of the difference between the true density function and the density function estimated from the data. The authors defined a general family of measures, called \(S\)-divergences, which includes many of the measures proposed previously as special cases. This paper gives a summary of the background and definition for the \(S\)-divergences and also various properties of the estimators, obtained by minimizing this type of divergence functions. It is shown that these estimators provide a high degree of stability under data contamination while attaining reasonable efficiency at the model and, in many cases, the most preferred estimators are obtained from the measures which have not been proposed previously. Some practical issues, such as tuning parameter selection and second order influence analysis, are also discussed.