A study of the influence of the ''natural restrictions'' on estimation problems in the singular Gauss-Markov model (Q1193813)

This paper considers the Gauss-Markov model \((Y,X\beta,\sigma^ 2 V)\). It is known that if the column space of the model matrix \(X\) is not contained in the column space of the dispersion matrix \(\sigma^ 2 V\), then the vector of parameters \(\beta\) has to satisfy certain linear equations. However these equations become restrictions on \(\beta\) in the usual sense only when the random vector \(Y\) involved in them is replaced by an observed outcome \(y\). The authors consider the problems of linear unbiased estimation and best linear unbiased estimation of an identifiable vector of parametric functions, comparison of estimators of any vector of parametric functions with respect to the matrix risk, and admissibility among the class of all linear estimators with respect to the matrix risk and with respect to the mean square error. Explicit solutions to these problems are derived when \(\beta\) is either unconstrained or constrained by the `natural restrictions' mentioned above. The conclusion established for every problem is that \(\beta\) may be considered as free to vary unconstrained without loss of generality, in the sense that if a given statistic is a solution to a given problem in the constrained space but is not a solution to the same problem in the unconstrained space, then there exists another statistic which is a solution to the problem in the unconstrained space.