Properties of modelling the error distribution with an extra shape parameter (Q1186056)

Based on asymptotic theory and some Monte Carlo simulations, it is shown that the adaptive stage of fitting an error distribution can as a first approximation be ignored in making statistical inference concerning the regression parameters provided the sample size is not small, and particulary if the \(T\) rather than the power-exponential distribution is used. This is a useful result which together with good robustness properties suggests that the adaptive robust techniques should have more of a role in practical statistics than they currently enjoy. However, caution should be exercised in extending this result to situations very different from those considered in the Monte Carlo study or to other more complex situations. For example the variance ratio for finite samples may not be so close to one for nonlinear, or multivariate models using the \(T\)-distribution or if the design of the explanatory variables is very irregular.