Robust Bayesian Method for Refutable Models

We propose a robust Bayesian method for economic models that can be rejected by some data distributions. The econometrician starts with a refutable structural assumption which can be written as the intersection of several assumptions. To avoid the assumption refutable, the econometrician first takes a stance on which assumption $j$ will be relaxed and considers a function $m_j$ that measures the deviation from the assumption $j$. She then specifies a set of prior beliefs $\Pi_s$ whose elements share the same marginal distribution $\pi_{m_j}$ which measures the likelihood of deviations from assumption $j$. Compared to the standard Bayesian method that specifies a single prior, the robust Bayesian method allows the econometrician to take a stance only on the likeliness of violation of assumption $j$ while leaving other features of the model unspecified. We show that many frequentist approaches to relax refutable assumptions are equivalent to particular choices of robust Bayesian prior sets, and thus we give a Bayesian interpretation to the frequentist methods. We use the local average treatment effect ($LATE$) in the potential outcome framework as the leading illustrating example.