Agricultural and Food Policy, Research Methods/Statistical Methods

The standard F test for linear restrictions in regression is relevant as a criterion but fails to capture the notion of tradeoff between bias and variance. Average squared distance criteria yield operational tests that are more appropriate, depending upon objectives. In the present paper two alternative criteria are developed. The first allows testing of the hypothesis that the average squared distance of a restricted estimator from the parameter point in k space is less than the average squared distance of the unrestricted, ordinary least squares estimator from the same parameter point. The second sets up a test of betterness of the restricted estimator over the unrestricted estimator of E(Y/X), where betterness is again defined in average squared distance.

Tables of critical points for the noncentral F are presented with noncentrality equal to 1/2 of numerator degrees of freedom for denominator degrees of freedom of 1-30, 40, 60, 120, 200, 400, and 1,000, and numerator degrees of freedom of 1-30, 40, 60, 120, and 200, and type one errors of 0.05, 0.10, 0.25, and 0.50. These critical points can be used to test the second weak MSE criterion discussed in the companion paper [2]. An approximation is suggested for noncentral F(θ), and accuracy checks are given. An appendix provides a Fortran function for the approximation. The approximation is intended for using the first weak MSE test discussed in the companion paper.