Title:A model-free approach to linear least squares regression with exact probabilities

Abstract:In a regression setting with observation vector $y \in R^n$ and given finite collection $(x_\nu)_{\nu \in N}$ of regressor vectors $x_\nu \in R^n$, a typical question is whether a given subset of these regressors is sufficient to approximate $y$. A classical method for this question is the F test, assuming that $y$ is a linear combination of the regressor vectors plus Gaussian white noise. In this note we show that the corresponding p-value has also a clear data-scientific interpretation without having to assume the data to be random. Then it is shown that such a dual interpretation is possible for a rather large family of tests, the underlying tool being normalized Haar measure on orthogonal groups.