The linear regression model: L<sub>p</sub>norm estimation and the choice of p
Arthur H. Money, J. F. Affleck‐Graves, Mike Hart +1 more
The LP norm provides alternatives to least squares for estimating the coefficients of a linear regression model. Monte Carlo studies were performed using six different values of p (including the least squares case of p =2) and were compared by means of the relative efficiencies based on the generalised variances. Values of p other than 2 showed improvement over least squares for all non-normal error distributions. A rule for determining a suitable p on the basis of the kurtosis of the error distribution is proposed.