Title: Bayesian Post-Selection Inference in the Linear Model

Abstract: We provide Bayesian inference for a linear model selected after observing the data. Adopting Yekutieli (2012)'s ideas, the Bayesian model consists of a prior and a truncated likelihood. The resulting posterior distribution, unlike in the setup usually considered when performing Bayesian variable selection, is affected by the very fact that selection was applied. After proposing a concrete extension of Yekutieli's framework to the case of variable selection, we turn to face the computational challenges associated with the adjusted posterior distribution. A major objection is the intractability of the truncated likelihood. At the core of our methods is a convex approximation to the truncated likelihood, which facilitates sampling from the (approximate) adjusted posterior distribution. We demonstrate both theoretically and in simulations that employing the proposed approximation results in Bayesian procedures that are qualitatively similar to those using the exact truncated likelihood.
Our methods are discussed in the context of recent literature on exact post-selection inference after model selection. These recent works focus on hypothesis testing, and capitalize on reductions achieved by conditioning out nuisance parameters. However, the techniques developed in that venue are generally less appropriate for addressing other questions, like point estimation. On the other hand, relying on an approximation to the full truncated likelihood, the tools we develop allow for more versatility. For example, replacing the genuine truncated likelihood by its approximation, we can approximate the maximum-likelihood estimate as the MAP estimate corresponding to a constant prior. We provide more examples in which our approximation can be employed to address frequentist questions that have not been resolved in existing work on exact post-selection inference.