Abstract : In estimating the coefficient of an endogenous variable in a single equation of a system of linear equations, Anderson and Sawa (1973) expressed the distribution of the two-stage least-squares (TSLS) estimator as a doubly noncentral F distribution. We relax their assumption of independent Gaussian errors, taking instead a scale mixture of spherical Gaussian laws in a class containing the spherical stable distributions. The resulting distribution of the TSLS estimator is a mixture of doubly noncentral F distributions mixed over the noncentrality parameters, and for suitable mixtures the normal-theory distribution is robust. Computations reported for contaminated spherical Gaussian and spherical Cauchy errors are compared with the standard case. (Author)