Overcoming model simplifications when quantifying predictive uncertainty