Missing data, in the form of gaps in the series of observations or caused by patient dropout, may occur frequently in repeated measures data (see Section 2.4.7). They are a less serious problem, however, when a mixed model is used, unless there are very substantial between-treatment group differences with respect to the pattern of dropout. The reason missing values are less problematical is that observations at each time point influence estimates of treatment effects at every other time point, owing to the specification of a covariance pattern. Thus, patients whose observations are limited to early time points because of dropout will nevertheless be taken into account when estimates are made of treatment effects at later time points. Clearly, such individuals will not influence these estimates as greatly as individuals whose data are complete, so the pattern of missing data in different treatment groups cannot be completely ignored. There will also be potential biases if patients show patterns of rapid deterioration prior to dropout. In such cases, their early observations maybe 'good' leading to a corresponding 'good' influence on the unobserved time points after dropout, when this is clearly inappropriate. Ad hoc approaches to imputing missing values, or analyses including and excluding dropouts, may then need to be employed. Alternatively, a method that allows for non-random missing data in repeated measures analysis could be considered (e.g. Diggle and Kenward, 1994). Thus, it is too simplistic to say that missing values do not matter in the mixed models analysis of repeated measures data, but the method is quite robust, even when the data may not be entirely missing at random.
Was this article helpful?