Resurrecting complete-case analysis: A defense

Complete-case analysis (CCA) is often criticized on the belief that CCA is only valid if data are missing completely at random (MCAR). Influential papers therefore recommend abandoning CCA in favor of methods that make a weaker missing-at-random (MAR) assumption. We argue for a different view: that CCA with principled covariate adjustment usefully complements MAR-based methods, such as multiple imputation. When estimating treatment effects, appropriate covariate control can, for some causal structures, eliminate bias in CCA. This can be true even when data are missing-not-at-random (MNAR) and when MAR-based methods are biased. We describe principles for choosing adjustment covariates for CCA, and we characterize the causal structures for which covariate adjustment does, or does not, eliminate bias. Even when CCA is biased, principled covariate adjustment often reduces the bias of CCA, and this method will sometimes be less biased than multiple imputation and other MAR-based methods. Therefore, when multiple imputation is used under the MAR assumption, adjusted CCA remains an important sensitivity analysis. When conducted with the same attention to covariate control that epidemiologists already afford to confounding, adjusted CCA belongs in the suite of reasonable methods for missing data. There is thus good justification for resurrecting CCA as a principled method.