On the optimality of finite Williams II(a) designs (Q1098524): Difference between revisions

\textit{R. M. Williams}, Biometrika 39, 151-167 (1952; Zbl 0046.360), considered experimental designs when the plots are contiguous and laid out in a line and when the errors or plot effects are assumed correlated according to a first or second order autoregressive process. For a first order autoregression AR(1) with positive lag-one correlation coefficient \(\lambda\), he looked at II(a) designs in which every treatment occurs equally often next to every other, but never to itself. There are some results on asymptotically optimum designs as n, the number of plots, tends to infinity. But, there are no known optimality results for finite n. \textit{D. R. Cox}, J. Roy. Statist. Soc., Ser. B 14, 211-229 (1952) and \textit{A. C. Atkinson}, Biometrika 56, 33-41 (1969) both conjectured A- optimality of II(a) designs for \(0<\lambda <1\). Here the investigation was done under the following restrictions: the type II(a) designs, having more than two treatments, with extra plot originally considered by Williams and the case of stationary AR(1) errors. In this paper, a number of interesting results are obtained. The authors show that Cox's and Atkinson's conjecture is true for all \(0<\lambda <1\) and can be extended to include D-optimality; an extension to E-optimality is not in general possible for all \(0<\lambda <1\), but is possible for \(\lambda\) not too small. For a restricted set of designs, a more general optimality of the II(a) designs for all \(0<\lambda <1\) is shown. The results are of importance in illustrating the problems that arise when autocorrelation is present and provide an interesting application of results in \textit{Ch. Cheng}, Ann. Stat. 15, 712-723 (1987; Zbl 0623.62069).