Robust and efficient estimation for the generalized Pareto distribution (Q2488456)

A minimum density power divergence estimator (MDPDE) is considered for the scale and shape parameters of the generalized Pareto distribution (GPD). Generally speaking, if the data \(X_i\), \(i=1,\dots,n\), are i.i.d. with PDF \(f(x;\vartheta)\), then the MDPDE for \(\vartheta\) is the minimizer of \[ \int f^{1+\alpha}(x;\vartheta)\,dx -(1+1/\alpha) n^{-1}\sum_{i=1}^n f^\alpha(X_i;\vartheta), \] where \(\alpha\) is a fixed positive number (the MLE corresponds to \(\alpha\to 0\)). The asymptotic variance and influence function (IF) of the MDPDE for the generalized Pareto distribution are described. It is shown that the IF is bounded for \(\alpha>0\) (in contrast to unbounded IF for MLE). The relative asymptotic efficiency of the MDPDE is investigated. The MDPDE is compared to the MLE, Hampel's optimally-biased robust estimator and the Peng-Welsh median estimator via simulations. An application to rainfall data is considered.