Systematic sampling for autocorrelated superpopulations (Q1263896): Difference between revisions

A study of systematic sampling is developed in the frame of a certain class of superpopulations. It is characterized by autocorrelation coefficients. The superpopulation model is the location one \((\theta +e_ iY_ i)\). The errors are random variables with zero mean and \(Cov(e_ i,e_ j)=\sigma^ 2\rho (| i-j|),\) where \(\rho\) (\(| i-j|)\) is an autocorrelation coefficient. This superpopulation approach uses predictive modelling and purposive sampling permits the selection of a sample without replacement for estimating the population total T. Two estimators are considered. One of them is obtained by searching for the BLUE (\^T\({}_ B)\) and the other by using the least squares principle (\^T\({}_ L)\). The design problem is fixed by the obtention of a sample and an estimator that minimizes the mean squared error (mse). Some features of \(\hat T_ L's\) mse are developed in Section 1. Section 2 considers that the units are arranged around a circle and that the condition \(\phi (i)=\phi (N-i)\) is imposed. Then variances and covariances do not depend on the design and \((n/N)^ 2V(\hat T_ L)\) should be minimized. A theorem states that equally spaced sampling (ess) is optimum if \[ P(i)=\rho (i)-2\rho (i+1)+\rho (i+2)\geq 0,\quad for\quad i\geq 0. \] Unfortunately there is not a similar result for \(\hat T_ B\). The arrangement in the line is tackled in Section 3. If \(N=2n\) and P(i) is decreasing ess is also optimum for \(\hat T_ L\). Conditions for a stationary process on R and stability assumptions are given and connected with the results of \textit{J. Kiefer} and \textit{H. P. Wynn} [Ann. Stat. 12, 431-450 (1984; Zbl 0558.62066)] and \textit{L. Shepp} [see Proc., Working Conf. Stoch. Proc. Sta. Barbara, CA.]. Under these assumptions an optimization problem can be solved.