When missingness depends on observed variables, the data available for training an imputation model can differ systematically from the population on which it is evaluated. The paper formulates imputation in terms of mean-squared error risk and uses importance weighting to account for this shift. Simulations compare the resulting imputations with unweighted baselines.