Missing at random

The requirement for an observation to be missing at random is that its expected value should be unaffected by whether or not the observation is missing. One of the most common reasons for missing values is patient non-compliance. If a patient decides that continued participation in a trial is just too much effort, or the patient dislikes the clinical procedures, and that decision is unrelated to any change in the outcome variables for the study, then the missing values will be missing at random. The above example is uncontroversial, but a more contentious example occurs if the patient withdraws from the trial because of a perceived adverse reaction to treatment. If the adverse reaction is unrelated to the outcome variables, then the missing data with respect to these efficacy variables can be regarded as missing at random. In taking this approach we are, in effect, attempting to estimate the effects of treatment on our outcome variables if adverse events do not occur. Of course, we do not ignore this important effect of treatment, and analysis of adverse events will form an important part in the evaluation of any clinical trial. By employing this philosophy in analysis, we are seeking to separate the effect of the treatments on the efficacy outcomes of interest from the assessment of the treatments' tolerability. It is probably the most usual way of dealing with patient withdrawal, but it will not always be the method of choice. For example, a pragmatic approach to a holistic evaluation of treatment would be to regard drug intolerability as a failure of treatment and predefine an appropriate value of the response variable for substitution when such a 'failure' occurs.

A decision as to whether missing values can be legitimately regarded as missing at random is rarely obvious. The advantages to the analysis if this assumption can be made are substantial, however, and so in practice this assumption is usually made unless there are strong grounds for concluding that the missing values are not missing at random.

Was this article helpful?

0 0

Post a comment