The purpose of this paper is three-fold. First, based on the asymptotic presentation of initial estimators, and model-independent parameters either hidden in the model or combined with the initial estimators, a pro forma linear regression between the initial estimators and the parameters is defined in an asymptotic sense. Then a weighted least squares estimation is constructed within this framework. Second, systematic studies are conducted to examine when both variance and bias reductions can be achieved simultaneously and when only variance can be reduced. Third, a generic rule of constructing composite estimation and unified theoretical properties are introduced. Some important examples such as quantile regression, nonparametric kernel estimation, blockwise empirical likelihood estimation are investigated in detail to explain the methodology and theory. Simulations are conducted to examine its performance in finite sample situations and a real dataset is analysed for illustration. The comparison with existing competitors is also made.