Comments on: Missing data methods in longitudinal studies: a review - Considerations for sensitivity analysis with likelihood-based models

摘要

Methods for analysis of incomplete data continue to play a prominent role in statistical and applied research. Joe Ibrahim and Geert Molenberghs are leading innovators and trend-setters in the field of missing data methodology, and have influenced both my own work and that of many others. The current paper makes a key contribution by collecting in one place a comprehensive and detailed account not only of methods for the analysis of incomplete longitudinal data, but their implementation. The contribution is significant in that it bridges a still appreciable gap between the desire and the ability to put advanced analytic methods into practice. Moreover, the paper lays out the full range of regression-based models that are available for handling incomplete data, covering both likelihood-based and moment-based estimation. A particularly important topic--and one that does not receive as much attention as it should--is missing covariates. In practice, missing covariates occur at least as frequently as missing responses. Ibrahim and Molenberghs have contributed a large proportion of the literature on missing data, and the current paper contains a welcome summary of key approaches.