Application of Stein-type estimation in combining regression estimates from replicated experiments (Q1342789)

This article considers the linear regression analysis of data arising from several experiments in which the responses are recorded for a set of preassigned values of the explanatory variables. Each data set thus provides estimates of regression coefficients in the model. An interesting question is then how to combine these estimates effectively in order to find an efficient estimate of the vector of regression coefficients. We have restricted our attention to the case of two experiments only for the sake of simplicity in handling the issue but our results can be extended without any considerable difficulty to the case of more than two experiments. The plane of our article is as follows. In section 2, we specify the model and describe the conventional procedure of taking a convex combination of the two least squares estimators of the coefficient vector. In order to find estimators that are more efficient than the convex combination, the application of Stein-type shrinkage procedure is explored and two families of estimators characterized by a single nonstochastic scalar are proposed. Section 3 discusses the properties of these estimators and presents a comparative study of the three estimators. Conditions for the superiority of one estimator over the other are obtained and therefrom some sufficient conditions are developed which do not involve any unknown parameter and are easily checked in any given application. Section 4 places some remarks and presents some suggestions. The Appendix provides the derivation of some results and some lemmas used within the text.