Consider the problem of estimating a <TEX>$p{\times}1$</TEX> mean vector <TEX>${\theta}$</TEX> (p <TEX>${\geq}$</TEX> 3) under the quadratic loss from multi-variate normal population. We find a James-Stein type estimator which shrinks towards the projection vectors when the underlying distribution is that of a variance mixture of normals. In this case, the norm <TEX>${\parallel}{\theta}-K{\theta}{\parallel}$</TEX> is known where K is a projection vector with rank(K) = q. The class of this type estimator is quite general to include the class of the estimators proposed by Merchand and Giri (1993). We can derive the class and obtain the optimal type estimator. Also, this research can be applied to the simple and multiple regression model in the case of rank(K) <TEX>${\geq}2$</TEX>.