5.14. Is MI better than single imputation of a draw from the predictive distribution of the missing values (SI) because (i) It yields more efficient estimates from the filled-in data? (ii) Unlike SI, it yields consistent estimates of quantities that are not linear in the data? (iii) It allows valid inferences from the filled-in data, if the imputation model is correct? (iv) It yields inferences that are robust against model misspecification?