Notes for a unified theory of estimation for sample surveys taking into account response errors (Q1393903)

In the introduction to the paper the author points out the need for putting together to pure theory of sample surveys and the theory of response and measurement errors which hitherto had developed almost along parallel lines. The definitions and notation given in section 2, needed for the development of a unified theory, may be considered to be a synthesis and perhaps a development of those already given by many authors working on the theory of estimation for sample surveys. (See references given in the paper.) The definitions on response and measurement errors and the associated probabilistic approach are generalizations of those proposed by \textit{M. H. Hansen}, \textit{W. N. Hurwitz} and \textit{M. Bershad} in their 1961 paper [``Measurement errors in censuses and surveys''. Bull. Int. Stat. Inst. 38, 359--374 (1961)]. The parameters due to response variations depend on the conditions under which the survey is carried out. In Section 3 the estimation of the uni-verse total is considered because it is the most basic problem. The author argues on the feasibility of taking into account initially all survey data in the construction of a linear estimator for the estimation of this total. In this estimator the coefficients attached to the variate values realized (which are subject to response and measurement errors) depend on the identity of the unit (unit label), the mark of order in the sequence of units realized by the particular sample, and the particular sample itself. Subsequently it is demonstrated that this estimator has larger mean square error than one based on the distinct sample with coefficients depending only the identity of the units and the distinct sample itself; further in the latter class of estimators it is shown that there is no estimator with the least mean square error, a result which is an analogue of Dodambe's result demonstrating the non-existence of the best linear unbiased estimator for estimation theory where observations are free from all errors. The unattained lower bound to this mean square error, which depends on the sampling design, the true values of the units for the characteristic in question and all the parameters associated with the response variances and covariances of all the units in the universe, is derived. It is also demonstrated that the information from entire sample data (defined along Fisherian lines) cannot be fully extracted. Section 4 shows that estimates of variance of all linear estimators in standard (textbook) theory are always underestimated. This is a very general result which applies to all sampling designs. The bias of underestimation can only be controlled by controlling the survey conditions. Finally section 5 describes the method of independent replication originated by Mahalanobis. There it is shown that variance estimation on the basis of estimates from the independent replicated samples can measure all the response variances and covariances except those induced by the replication procedure itself.