On frequentist coverage errors of Bayesian credible sets in moderately high dimensions (Q2278674)

Let \(\mathbf{Y}=\mathbf{X}\bar{\beta}+\mathbf{r}+\bar{\varepsilon}\) be a linear regression model, where \(\mathbf{Y}=(Y_1,\ldots,Y_n)^\intercal\) is a vector of outcomes, \(X\) is an \(n\times p\) design matrix, \(\bar{\beta}=(\beta_1,\ldots,\beta_p)\) is an unknown coefficient vector, \(\mathbf{r}=(r_1,\ldots,r_n)^\intercal\) is a deterministic bias term, and \(\bar{\varepsilon}=(\varepsilon_1,\ldots,\varepsilon_n)^\intercal\) is a vector of i.i.d. error terms with zero means and finite variance \(\sigma^2\). It is supposed that the number of regressors \(p\) increases together with the sample size \(n\), and it is allowed the error variance \(\sigma^2\) to depend on \(n\). It is supposed also that prior distributions \(\Pi_{\bar{\beta}}\), \(\Pi_{\sigma^2}\) of \(\bar{\beta}\), \(\sigma^2\) are independent, and \(\Pi_{\bar{\beta}}\) is absolutely continuous. Let \(\widehat{\beta}=(\mathbf{X}^\intercal \mathbf{X})^{-1}\mathbf{X}^\intercal \mathbf{Y}\) denote the \textit{ordinary least squares} estimator for \(\beta\) with \(\mathbf{r}=\mathbf{0}^\intercal\). For a given \(\alpha\in(0,1)\), let \(I(\widehat{\beta},\widehat{R}_\alpha)\) be a \((1-\alpha)\)-credible rectangle, where the radius \(\widehat{R}_\alpha\) is chosen in such a way that \(\Pi_{\bar{\beta}}\{I(\widehat{\beta},\widehat{R}_\alpha)\,|\,\mathbf{Y}\}=1-\alpha\). The authors obtain an upper bound for the quantity \[ \left| \mathbb{P}\left(\bar{\beta}\in I(\widehat{\beta},\widehat{R}_\alpha)\right)-(1-\alpha)\right| \] provided for every \(n\geqslant 2\) under relatively loose restrictions on the components of the model.