When Can Multiple Imputation Improve Regression Estimates?

Abstract

Multiple imputation (MI) is often presented as an improvement over listwise deletion (LWD) for regression estimation in the presence of missing data. Against a common view, we demonstrate anew that the complete case estimator can be unbiased, even if data are not missing completely at random. As long as the analyst can control for the determinants of missingness, MI offers no benefit over LWD for bias reduction in regression analysis. We highlight the conditions under which MI is most likely to improve the accuracy and precision of regression results, and develop concrete guidelines that researchers can adopt to increase transparency and promote confidence in their results. While MI remains a useful approach in certain contexts, it is no panacea, and access to imputation software does not absolve researchers of their responsibility to know the data.