Abstract
We consider the linear calibration model: $y_1={\alpha}+{\beta}x_i+{\sigma}{\varepsilon}_i$, i = 1, ${\cdots}$, n, $y={\alpha}+{\beta}x+{\sigma}{\varepsilon}$ where ($y_1$, ${\cdots}$, $y_n$, y) stands for an observation vector, {$x_i$} fixed design vector, (${\alpha}$, ${\beta}$) vector of regression parameters, x unknown true value of interest and {${\varepsilon}_i$}, ${\varepsilon}$ are mutually uncorrelated measurement errors with zero mean and unit variance but otherwise unknown distributions. On the basis of simple small-sample low-noise approximation, we introduce a new method of comparing the mean squared errors of the various competing estimators of the true value x for finite sample size n. Then we show that a class of estimators including the classical and the inverse estimators are consistent and first-order efficient within the class of all regular consistent estimators irrespective of type of measurement errors.