ABSTRACT
There are numerous methods of imputation that are becoming easier to perform with existing software. As a result, there is a danger of using imputation as a quick fix for problems that chould have been prevented with a thoughtful trial design. As we will illustrate in this and the next chapter, all of the methods rely on strong assumptions about the missing data mechanism which are often ignored. In most cases when the assumptions are correct there are direct methods of analysis utilizing all of the available data that produce the same results (see Chapters 3 and 4). If the concern is that the missingness is non-random (MNAR), imputation methods will only be helpful if there is additional auxiliary information. The following quote summarizes the issue:
It is clear that if the imputation model is seriously flawed in terms of capturing the missing data mechanism, then so is any analysis based on such imputation. This problem can be avoided by carefully investigating each specific application, by making the best use of knowledge and data about the missing-data mechanism, ... [Barnard and Meng, 1999]
There are a limited number of situations where imputation is useful in HRQoL research. The first occurs when the respondent skips a small number of items in a questionnaire (Section 8.2). The second occurs when there are multiple explanatory (independent) variables or covariates with very small proportions of missing data individually, but a substantial proportion overall [vanBuuren et al., 1999] (Section 8.5). Missing covariates will result in the deletion of individuals with missing data from any analysis. The concern is that a selection bias occurs when those with incomplete data differ from those with complete data. Missing assessments (entire questionnaire) are a more common problem in HRQoL research. Only when there is auxiliary information such as surrogate variables or strongly correlated clinical data will imputation methods be useful. In Sections 8.3 and 8.4, I will focus on imputation of the HRQoL measures, not because I would recommend simple imputation methods for missing outcome measures, but because it is easier to illustrate the methods and principles.
