On the prediction of means in a finite population (Q1079306)

Consider a general superpopulation model with \[ Y_ t=g(X_ t;d_ 0)+e_ t \] where \(E(e_ t| X_ t)=0\), \(V(e_ t| X_ t)=\sigma^ 2a_{tt}\) and \(Cov(e_ t,e_ s| X_ t,X_ s)=0\) and g has a known parametric form and is continuous. When a consistent estimator of \(d_ 0\), say \(\hat d_ 0\) is obtained, consider the predictor \[ \bar y_ A=N^{-1}n\bar y_ n+(1-N^{-1}n)(N-n)^{- 1}\sum^{N}_{t=n+1}g(X_ t;\hat d_ 0) \] where \(\bar y_ n\) is the sample mean. Alternatively one may consider the predictor \[ \bar y_ B=N^{-1}\sum^{N}_{t=1}g(X_ t;\hat d_ 0). \] The author compares \(\bar y_ A\) and \(\bar y_ B\) under the above model, and obtains necessary and sufficient conditions for superiority of \(\bar y_ B\) over \(\bar y_ A\) in large samples. When weighted least squares method is used to estimate \(d_ 0\) with weights inversely proportional to \(a_{tt}\), it is proved that \(\bar y_ A\) is asymptotically more efficient than \(\bar y_ B\). The effects of measurement errors on the relative performance of the two predictors are also examined.