Likelihood based confidence intervals for the tail index (Q1424687)

Let \(X_{1},\ldots,X_{n}\) be i.i.d. observations with distribution function \(F\) satisfying \(\lim_{t\to\infty} [1-F(tx)]/[1-F(t)]= x^{-1/\gamma}\), for all \(x>0\), where the parameter \(1/\gamma>0\) is called tail index. This paper investigates two methods, namely the empirical likelihood method and a parametric likelihood method, for constructing confidence intervals of the tail index. Let us define \(Y\) as the ratio of natural logarithms of the \((n-i+1)\)-th and \((n-i)\)-th order statistics given as follows: \(Y_{i}= i\log(X_{n,n-i+1}/ X_{n,n-i})\) for \(i=1,\ldots,k\). Denote by \(U(x)\) the inverse function of \(1/(1-F(x))\). A nonparametric confidence interval for the tail index can be constructed based on the following result. Let there exist a function \(A(t)\to0\), as \(t\to\infty\), such that \[ A^{-1}(t) [U(tx)/U(x)-x^{\gamma}]= x^{\gamma} \rho^{-1} (x^{\rho}-1) \quad \text{for all }x>0, \] where \(\rho<0\), and \(k\to\infty\), \(k/n\to0\) as \(n\to\infty\). Then \(l(\gamma)= 2\sum_{i=1} ^{k}\log\{1+\lambda(Y_{i}-\gamma)\}\), where \(\lambda\) is determined by \(\sum_{i=1} ^{k}[Y_{i}- \gamma]/[1+\lambda (Y_{i}-\gamma)]=0\), has an asymptotic \(\chi^{2}_{1}\) distribution if \(\sqrt{k}A(n/k)\to0\) as \(n\to\infty\). Under the same condition on \(U(x), k, n\), \(R(\gamma)= 2(k+1)[\overline \gamma _{n}/ \gamma-1-\log (\overline\gamma_{n}/\gamma)]\) has an asymptotic \(\chi_{1}^{2}\) distribution provided \(\sqrt{k}A(n/k) \to0\) as \(n\to\infty\), where \[ \overline\gamma_{n}= (k+1)z ^{-1} \sum_{i=1}^{k} \log(X_{n,n-i+1}/X_{n,n-k}). \] A given simulation study compares the empirical likelihood method, the parametric likelihood method and the normal approximation method in terms of convergence in probability.