Comparison of linear experiments with known covariances (Q797928): Difference between revisions

A linear experiment with known covariances, denoted by \(L(X\beta\),V) is represented by the model \(Y=X\beta +\epsilon\), \(E(\epsilon)=0\), \(Var(\epsilon)=V\), where V may be singular. For two linear experiments \(d_ 1=L(X_ 1\beta,V_ 1)\) and \(d_ 2=L(X_ 2\beta,V_ 2)\), where \(V_ 1,V_ 2\) are known, \textit{S. Ehrenfeld} [Proc. 3rd Berkeley Sympos. math. Statist. Probability 1, 57-67 (1956; Zbl 0075.147)] defined \(d_ 1\) to be at least as good as \(d_ 2 (d_ 1\geq d_ 2)\) if and only if every c'\(\beta\) which is estimable in \(d_ 2\) is also estimable in \(d_ 1\) and var(c'\({\hat\beta }{}_ 1)\leq var(c'{\hat\beta }_ 2)\) for all such c', where \({\hat\beta }{}_ i\) is the best linear unbiased estimator of \(\beta\) in \(d_ i\), \(i=1,2.\) The authors obtain necessary and sufficient conditions for the relations \(d_ 1\geq d_ 2\) to hold. They define the relation \(''d_ 1\) strictly better than \(d_ 2'' (d_ 1>d_ 2)\) and obtain necessary and sufficient conditions for this to hold. Furthermore they extend all these results to the case where only a subset of the parameters is of interest.