Uncertain identification

Uncertainty about the choice of identifying assumptions is common in causal studies, but often ignored empirical practice. This paper considers uncertainty over models that impose different assumptions, which can lead to a mix point‐ and set‐identified models. We propose performing inference presence such by generalizing Bayesian model averaging. The method multiple posteriors for combines them with single posterior are either point‐identified or nondogmatic assumptions. output set ( post‐averaging ambiguous belief ), be summarized reporting means associated credible region. clarify when prior probabilities updated characterize asymptotic behavior probabilities. provides formal framework conducting sensitivity analysis findings For example, we find standard monetary one would need attach probability greater than 0.28 validity assumption prices do not react contemporaneously policy shock, order obtain negative response shock.