Multiple Imputation of Missing Values: Update of Ice

TLDR

Royston introduced mvis, a Stata implementation of MICE for multiple multivariate imputation under MAR, and later upgraded it to ice incorporating user‑driven improvements. This article updates ice, explaining the modest but important changes. The update changes ice’s default imputation to sample from the posterior predictive distribution rather than predicted mean matching, while retaining the five ado‑files (ice, micombine, mijoin, misplit, uvis) with the last three unchanged.

Abstract

Royston (2004) introduced mvis, an implementation for Stata of MICE, a method of multiple multivariate imputation of missing values under missing-at-random (MAR) assumptions. In a second article, Royston (2005) described ice, an upgrade incorporating various improvements and changes to the software based on personal experience, discussion with colleagues, and user requests. This article describes an update to ice. The changes are less substantial but nevertheless important enough to warrant a brief explanation. The major modification is that the default method of imputing missing values in ice is now by sampling from the posterior predictive distribution rather than by predicted mean matching. The ice system comprises five ado-files: ice, micombine, mijoin, misplit, and uvis. The last three programs have not been changed and are included in the present release for the sake of completeness.