Exact linear unbiased estimation in survey sampling

The concept of the linearity of estimators in finite population inference is not well defined. We propose that linearity should be defined for both estimators and estimands and that strict linearity should be related constructively to the property of unbiasedness. We establish the conditions for the class of general linear estimators, introduced by Godambe (J. Roy. Statist. Soc. 17 (1955) 269), to be strictly linear. The ideas are extended to quadratic estimators of quadratic estimands. The implications are explored for ratio estimators, and domains of study, the latter extendible to poststratification