In this video, we’re looking at what multiple imputations are and how they can be used to deal with missing data.
I explain why missing data can be a source of bias, using a very simple dataset to illustrate the problems. The default approach in statistical packages is to remove cases with missing values, so-called list-wise deletion. This is not necessarily a bad thing, but normally, we don’t really know why data are missing, so we don’t know if and how much bias we have.
We look at replacing missing value with the mean and using separate categories, and see that these (common) approaches are inadequate. Multiple imputations are preferable because they keep information about how uncertain we are about the imputed data. This is done by drawing from a distribution and running analyses multiple times before combining the results.
The video doesn’t go into the technicalities of multiple imputations and the different approaches there are to implement the basis idea in practice, but it should be clear that even multiple imputations cannot do magic.
Didier Ruedin 