Exact adaptive confidence intervals for linear regression coefficients (Q131754)

In this paper, the authors consider an alternative to the uniformly most accurate unbiased confidence interval procedure for the coefficients in the normal linear regression model \(\mathbf{y} \sim N_n(X\mathbf{\beta}, \sigma^2)\). They present an adaptive confidence interval procedure for adaptively estimating a normal prior distribution for \(\mathbf{\beta}\) from the data \(\mathbf{y}\) to reduce the potential risk of a poorly specified prior, and then using this estimated prior distribution to construct an approximately Bayes-optimal confidence interval procedure for each regression coefficient \(\beta_j,\) \(j = 1, \ldots, p\). The resulting adaptive confidence intervals maintain exact non-asymptotic \(1-\alpha\) coverage if the design matrix is full rank and the errors are normally distributed. Note that this procedure has a frequentist coverage rate that is constant as a function of the model parameters, yet provides smaller intervals than the usual interval procedure, on average across regression coefficients. No assumptions on the unknown parameters are necessary to maintain exact coverage. Additionally, in a ``\(p\) growing with \(n\)'' asymptotic scenario, this adaptive ``frequentist, assisted by Bayes'' procedure is asymptotically Bayes-optimal among \(1-\alpha\) frequentist confidence interval procedures.