A number of high-profile epidemiologic papers utilize imputation to "fill the gaps" in partially incomplete data. Should this be encouraged or discouraged? Does imputation increase or decrease your "trust" in reported results? Should there be standardized rules regarding this practice?
There are many forms of imputation. If you mean 'multiple imputation', that method depends on the assumption that data is "missing at random (MAR)". That means that the data is missing due patterns in some measured covariate in your dataset. This can rarely be verified in observational studies. As long as: 1) your dataset is large enough, 2) you correctly specify the imputation model with the covariates which are informative to the missing data, 3) data is MAR, and 4) you perform enough rounds of imputation, the method is fine and you should maintain statistical efficiency with missing data without introducing substantial bias.
However, multiple imputation can't really be used with repeated-measures / longitudinal data. In these cases, most people use 'last observation carried forward' which is a special case of imputation. However, the most feasible and valid way to handle missing longitudinal data is using 'inverse probability time-varying weights' assuming data is MAR.
- Badges
- Science topic
- Similar topics
- Papers