An ancillarity paradox which appears in multiple linear regression (Q2641018)

Consider a multiple regression in which \(Y_ i\), \(i=1,...,n\), are independent normal variables with variance \(\sigma^ 2\) and E \(Y_ i=\alpha +V'_ i\beta\) where \(V_ i\in {\mathbb{R}}^ r\) and \(\beta \in {\mathbb{R}}^ r\). Let \({\hat \alpha}\) be the usual least squares estimator of \(\alpha\). If \(V_ i\) is fixed than \({\hat \alpha}\) is admissible under squared error loss. However, if \(V_ i\) are random observations, independently distributed according to a multivariate normal distribution with mean \(\underset{\tilde{}} 0\) and known nonsingular covariance matrix \(\theta\), then (it is shown that) \({\hat \alpha}\) is inadmissible, when \(r\geq 2\). Several estimators dominating \({\hat \sigma}\) when \(r\geq 3\) are given. Analogous results are given for the case when \(\theta\) and \(\sigma^ 2\) are unknown. It is interesting to note that \(\{V_ i\}\) are ancillary statistics in the above setting. Hence the admissibility of \({\hat \alpha}\) depends on the distribution of the ancillary statistics. This fact constradicts the notion, widely held, that statistical inferences should be carried out, conditional on the value of any ancillary statistic. That is, the distribution of the ancillary statistic should be irrelevant to the statistical inference.