Calibration with a nonlinear standard curve: how to do it? (Q1965957)

A parameter \(z_0\in(a,b)\subseteq R\) is estimated from observations \(Z_j=f(z_0,\theta_0)+\eta_j\), \(j=1,\dots,m\), where \(\eta_j\) are i.i.d. errors, \(f(\cdot,\theta)\) is a known continuous, strictly monotone function, \(\theta_0\) is an unknown parameter, \(\theta_0\in\Theta\subset R^p\), and \(\Theta\) is compact. The true value of \(\theta_0\) is estimated by means of a complementary (calibration) experiment in which \(n\) random variables \(X_i\) are observed, such that \(X_j=f(x_j,\theta_0)+\varepsilon_i,\) \(i=1,\dots,n,\) where \(x_i\) are known values and \(\varepsilon_i\) are i.i.d. errors. The authors consider five asymptotic confidence sets for \(z_0\). Two of them are based on the least squares estimator \(\hat\theta_n\) for \(\theta\), \(\bar Z_m = m^{-1}\sum_{j=1}^m Z_j\), and the estimator \(\hat z_{m,n}=g(\bar Z_m,\hat\theta_n)\) for \(z_0\), where \(g(\cdot,\theta)=f^{-1}(\cdot,\theta)\). Three other sets are based on the difference \(\bar Z_m - f(z,\hat\theta_n)\). Edgeworth expansions are used to determine the asymptotics of confidence levels for these sets as \(n\to\infty\), \(m=n^{\delta}\), \(0<\delta<4/5\). Bootstrap modifications of the confidence sets are proposed and their consistency is proved under the same asymptotics. The results of simulations are presented for the model \(f(z,\theta)=\theta_1(1+\exp(\theta_2+\theta_3 z)^{-1}\) with \(\varepsilon\) and \(\eta\) being gamma distributed.