A minimax approach to Bayesian estimation with partial knowledge of the observation model