scientific article; zbMATH DE number 4015957
[For the entire collection see Zbl 0583.00027)] We consider the gaussian linear model (1) \(Y=X\beta +\epsilon\) where \({\mathcal L}(Y)=N_ n(\mu,\sigma^ 2I_ n)\), \(\sigma >0\), \(\mu\in Q\) (a linear subspace of \({\mathbb{R}}^ n)\). This model is invariant under the group of transformations \(Y\to aY+\xi (a>0\), \(\xi\in Q)\) and a maximal invariant is the vector T of normed residuals. Thus, if (1) is considered as a null hypothesis to be challenged, the restriction to invariant procedures leads to perform the analysis via T. The matrix approach is not very convenient to deal with T because the mapping \(Y\to T\) is not linear. In fact, if model (1) is true, T is uniformly distributed on the unit sphere S of \(Q^{\perp}\). Under some alternatives it is easy to compute the density of T with respect to the uniform probability on S, for example when \({\mathcal L}(Y)=N_ n(\nu,W)\), \(\nu\not\in Q.\) The formulation which we advocate leads to straightforward results concerning the optimality of some procedures and enables us to give a clear account of the assumptions which are used. The problem of k outliers detection is more specially discussed, including the case where k is not fixed.
This page was built for publication:
Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q3028121)