Robust estimation of tail parameters for two-parameter Pareto and exponential models via generalized quantile statistics (Q5954056)

The authors consider an i.i.d. sample \(X_1\), \dots, \(X_n\) with d.f. \(F(x)=1-(\sigma/x)^\alpha\) (\(x>\sigma\)) or, equivalently, \(Z_1\),\dots, \(Z_n\) with d.f. \(G(x)=1-\exp(-(z-\mu)/\vartheta)\) (\(z\geq\mu\)). (Note, that \(\ln X_j\) has d.f. \(G\) with \(\mu=\log\sigma\), \(\vartheta=\alpha^{-1}\)). The MLE of \(\vartheta\) (\(\mu\) being unknown) is \(\hat \vartheta_{ML}=n^{-1}\sum_j Z_j-\min_j\{Z_j\}\). Evidently this estimator is very nonrobust. The authors discuss many other estimators (trimmed mean; least squares; linear model based; generalized quantile). These estimators are compared in asymptotic relative efficiency w.r.t. \(\hat\vartheta_{ML}\) and breakdown point as a robustness criterion. The authors' conclusion is that their generalized median estimator \(\hat \vartheta_{GM}\) ``dominates the other competitors and should become incorporated into practice''. The estimator \(\hat\vartheta_{ML}\) is defined as follows. Let \[ h(z_1,\dots,z_k)=k^{-1}\sum_{j=1}^k z_j-\min\{z_1,\dots, z_k\},\quad \hat H_n(y)=\binom{n}{k}\sum \{h(Z_{i_1},\dots,Z_{i_k})<y\}, \] where the sum is taken over all \(k\)-sets of distinct indices \((i_1,\dots,i_k).\) Then \(\hat \vartheta_{GM}=H_G^{-1}(1/2)\).