On Expected Lengths of Predictive Intervals (Q4828217)

Let \(X_0\) be a scalar-valued observation with PDF \(f(x;\vartheta)\) and CDF \(F(x;\vartheta)\), where \(\vartheta\) is an unknown parameter. The authors consider a predictive interval \(S=[a(\vartheta), b(\vartheta)]\) for \(X_0\) such that \(f(a(\vartheta),\vartheta)=f(b(\vartheta),\vartheta)\) and \(\Pr_\vartheta(X_0\in S)=F(b(\vartheta),\vartheta)-F(a(\vartheta),\vartheta)=1-\alpha\), where \(\alpha\) is fixed. Since \(\vartheta\) is unknown it is replaced by an estimator \(\widehat\vartheta_n\) from an i.i.d. sample \(X_1\),\dots, \(X_n\) with PDF \(f\). The approximate predictive interval is \[ S_n=[a(\widehat\vartheta)+n^{-1}a_1(\widehat\vartheta)+n^{-2}a_2(\widehat\vartheta), b(\widehat\vartheta)+n^{-1}b_1(\widehat\vartheta)+n^{-2}b_2(\widehat\vartheta)], \] where \(a_i\), \(b_i\) are some smooth functions to be chosen such that \(\Pr_\vartheta(X_0\in S)=1-\alpha+o(n^{-2})\). An asymptotic expansion is derived for the expected length \(l\) of \(S_n\) in the form \[ E_\vartheta l(S_n)=c_0+n^{-1}\Omega_1+n^{-2} (\Omega_2+\omega^{-1}\Delta)+o(n^{-2}), \] where \(c_0\), \(\Omega_1\), \(\Omega_2\) and \(\omega\) are functions of \(\vartheta\) that do not depend on \(n\), \(a_i\) and \(b_i\). The problem of optimal choice of \(a_i\) and \(b_i\) is considered. Second order admissibility of the intervals is discussed.