Calibration of predictive distributions in imputation methods

Determine whether the predictive distributions and prediction intervals produced by imputation methods that output uncertainty are calibrated in the sense that their nominal probabilities match empirical frequencies.

Background

The paper argues that prior work on missing-data imputation has focused primarily on point accuracy and has provided only informal or incomplete treatment of uncertainty quality. As a result, the reliability of uncertainty estimates remains insufficiently characterized across different methods and settings.

The authors emphasize the importance of calibration—whether nominal coverage agrees with empirical coverage—and note that this has not been systematically established for widely used approaches, motivating their empirical study across multiple datasets and missingness mechanisms.