Abstract:
Artificial linear regressions often provide a convenient way to calculate test statistics and estimate covariance matrices. This paper discusses one family of these regressions, called "double-length" because the number of observations in the artificial regression is twice the actual number of observations. These double-length regressions can be useful in a wide variety of situations. They are easy to calculate, and seem to have good properties when applied to samples of modest size. We first discuss how they are related to Gauss-Newton and squared-residuals regressions for nonlinear models, and then show how they may be used to test for functional form and other applications.
