The trimmed mean in the linear model (Q1822433)

Let \(X=Ba+e\) be a general linear model, i.e. B is a known \(n\times p\)- matrix, \(a\in R^ p\) is an unknown parameter vector and e is an n- dimensional random vector with i.i.d. components. The author proposes a generalization \(\tau_ n\) of trimmed means to estimate \(a\) which has similar asymptotic properties as the corresponding estimator in the location model, if one imposes suitable conditions on a preliminary estimator \(a_ n\) for \(a\), on the design matrix B and the distribution of e. The definition of \(\tau_ n\) is based on the von Mises functional representation of trimmed means in the location model. Discussed is also a version of \(\tau_ n\) where the trimming proportions are based on the data.