Estimating model prediction error: Should you treat predictions as fixed or random? uri icon