This is an archived version of the Handbook. For the current version, please go to training.cochrane.org/handbook/current or search for this chapter here.

A common, but potentially dangerous, approach to dealing with missing outcome data is to **impute** outcomes and treat them as if they were real measurements (see also Chapter 16, Section 16.2). For example, individuals with missing outcome data might be assigned the mean outcome for their intervention group, or be assigned a treatment success or failure. Such procedures can lead both to serious bias and to confidence intervals that are too narrow. A variant of this, the validity of which is more difficult to assess, is the use of ‘last observation carried forward’ (LOCF). Here, the most recently observed outcome measure is assumed to hold for all subsequent outcome assessment times (Lachin 2000, Unnebrink 2001). LOCF procedures can also lead to serious bias. For example, in a trial of a drug for a degenerative condition, such as Alzheimer’s disease, attrition may be related to side effects of the drug. Because outcomes tend to deteriorate with time, using LOCF will bias the effect estimate in favour of the drug. On the other hand, use of LOCF might be appropriate if most people for whom outcomes are carried forward had a genuine measurement relatively recently.

There is a substantial literature on statistical methods that deal with missing data in a valid manner: see Chapter 16 (Section 16.1). There are relatively few practical applications of these methods in clinical trial reports (Wood 2004). Statistical advice is recommended if review authors encounter their use. A good starting point for learning about them is www.missingdata.org.uk.