THE EXACT MEAN SQUARED ERROR OF STEIN-RULE ESTIMATOR IN LINEAR MODELS Shyamal Das PEDDADA Dept. of Mathematics, Central Michigan University, Mt. Pleasant, MI, and Dept. of Mathematics/Statistics, University of Nebraska, Lincoln, NE, USA
Parthasarathi
LAHIRI
Dept. of Mathematics/Statistics,
Received
16 January
Recommended
University of Nebraska, Lincoln, NE, USA
1987; revised
by M.L.
Abstract: The exact Mean Squared tion of the regression estimator
within
manuscript
received
23 April
1987
Puri
vector
Error
is obtained.
an ellipsoid
(MSE) of the Stein Rule estimator Further,
the inadmissibility
for a linear combina-
of the usual least squares
is investigated.
AMS Subject Classification: 62507. Key words and phrases: Ellipsoid;
fractional;
linear model;
non-normal;
operators.
1. Introduction
In the usual linear model Y=Xfl+ E, Y is an 12x 1 vector of observations, X is an n x m design matrix, p is an m x 1 vector of unknown parameters and E is an n x 1 random error vector. We shall assume that E has a multivariate normal distribution with mean vector 0 and covariance matrix 02Z, where o2 is unknown. Without loss of any generality one can assume that X’X=nZ. Under this model the usual least squares estimator for o2 is s2 which is proportional to Y’[Z- n-‘XX’] Y. A well known modification to the estimator, more commonly known as the Stein-rule estimator, is bi = (1 - as2/nb’b)b, where a is a scalar constant. It has been established that under squared error loss, 6i dominates b if 0
0