Saddlepoint approximations for the distribution of some robust estimators of the variogram (Q2303033): Difference between revisions

Let a random variable \(Z_s\) be observed at some fixed subset of \(\mathbb{R}^d\), \(d \geq 1\). In Geostatistics, it is important to know the distribution of the variogram \(2\gamma(h)=E(Z_s-Z_{s+h})^2\). Although the total number of observations can be large, the sample size at a given point \(s\) is often small. In the article, accurate robust saddlepoint approximations of a variogram are obtained in the case of small-sample sizes. As some small-sample approximations are known only if \((Z_s)_i\) are i.i.d. Gaussian random variables, the scale contaminated normal model is considered. According to this model, \(Z_s\) is described by the PDF \(F=(1-\epsilon)N(\mu,\sigma)+\epsilon N(\mu,g\sigma)\) with \(\epsilon \in (0,1)\) (usually small) and \(g>1\). This makes it possible to use von Mises approximations. In the article classical, Matheron's estimator, the robust Huber's variogram estimator and the \(\alpha\)-trimmed variogram estimator are obtained. To obtain the approximations, original observations are transformed into a new ones, which can be considered independent if the condition of a linear semivariogram model \(\gamma(2h)=2\gamma(h)\) holds. To check this, a goodness of fit test is defined.