Mixed models are much more flexible than fixed effects models in the treatment of missing values. For example, in a two-period, cross-over trial, information on subjects with one value missing is completely lost when a fixed effects analysis is used. In contrast, mixed models are capable of handling the imbalance caused by missing observations provided that they are missing at random. This is often a reasonable assumption to make. If a subject withdraws from a cross-over trial after receiving one treatment, then we may have no idea of how the subject would have responded to the other treatments, and to handle these non-observed periods as if they were missing at random seems eminently sensible. It is helpful at this stage to classify missing data into one of three widely used categories.
Was this article helpful?