Semiparametric inference for a quantile comparison function with applications to receiver operating characteristic curves (Q4935353)

The authors consider the independent, two-sample random censoring model where one population is modeled parametrically and the other non-parametrically. Let \(F\) denote the nonparametric cumulative density function (cdf) of the first sample and \(G_\theta\) denote the parametric cdf of the second sample with \(k\)-dimensional parameter \(\theta\) and with respective censoring cdfs \(H_1\) and \(H_2\). The vertical quantile comparison function \(Q(p)= G_\theta(F^{-1}(p))\), \(0< p< 1\), is used for semiparametric two-sample inference. Applying standard counting process techniques the asymptotic distribution of the semiparametric estimator \(Q_{\widehat\theta}(p)= G_{\widehat\theta}(\widehat F^{-1}(p))\) is obtained if \(\widehat\theta\) is the MLE of \(\theta\). This estimator is compared with the complete nonparametric estimator \(\widehat\theta(p)=\widehat G(\widehat F^{-1}(p))\) for its asymptotic relative efficiency. It is shown that the asymptotic variance of \(Q_{\widehat\theta}(p)\) can be substantially smaller than that of \(\widehat Q(p)\), especially when \(p\) is near \(0\) or \(1\). The limiting distribution of \(W_m(p)= \sqrt m\{Q_{\widehat\theta}(p)- Q_{\theta}(p)\}\) is investigated using the Khmaladze transform as well as using a bootstrap approximation, chosen nonparametrically for \(F\) and semiparametrically for \(G_\theta\). Using this distribution theory, Kolmogorov-Smirnov-type and chi-squared tests are constructed for the hypotheses \(H_0: F=G_\theta\) based on \(Q_{\widehat\theta}(p)= Q(p)\). Finally, confidence bands and two-sample tests are obtained for \(\text{ROC}_{\widehat\theta}(p)= 1-Q_{\widehat\theta}(1- p)\) as an estimate of the receiver operating characteristic (ROC) curve \(\text{ROC}(p)= 1-Q(1- p)\). Related results are given for the areas under the ROC curve.