Confidence Intervals for Bisquare Regression Estimates

This paper describes the results from a Monte Carlo study of robust regression confidence-interval estimation in the model y = a + bx + ε. Bisquare estimators were studied on samples of 11 and 21 with five design matrices. Four estimators of scale were used to form confidence intervals. Samples were generated from two distributions: normal and slash. For many design matrices, the four scale estimators give almost equally efficient confidence intervals but require different t values to achieve this efficiency. However, for each scale estimator, the t value changes little as a function of which linear combination of the parameters is being estimated.