Median based covariogram estimators reduce bias (Q1062403)

The usual product moment covariogram estimator of a Gaussian process can have appreciable bias. This article shows that the bias is decreased when the sample mean \(\bar X\) in the estimator is replaced with med(X), the sample median (provided a mild condition on the negative covariances is satisfied). Identical conclusions are drawn, even when the Gaussian process is (symmetrically) contaminated.