It is examined how Bayesian inference might proceed in models with many instruments. A new prior specification based on Lancaster's (1997) parameter orthogonalization is developed. This orthogonalization is shown to guarantee that the statistical problems associated with the first stage...