In full ITT analyses, all participants who did not receive the assigned intervention according to the protocol as well as those who were lost to follow-up are included in the analysis. Inclusion of these in an analysis requires that means and standard deviations of the outcome for all randomized participants are available. As for dichotomous data, dropout rates should always be collected and reported in a ‘Risk of bias’ table. Again, there are two basic options, and in either case a sensitivity analysis should be performed (see Chapter 9, Section 9.7).
Available case analysis: Include data only on those whose results are known. The potential impact of the missing data on the results should be considered in the interpretation of the results of the review. This will depend on the degree of ‘missingness’, the pooled estimate of the treatment effect and the variability of the outcomes. Variation in the degree of missing data may also be considered as a potential source of heterogeneity.
ITT analysis using imputation: Base an analysis on the total number of randomized participants, irrespective of how the original study authors analysed the data. This will involve imputing outcomes for the missing participants. Approaches to imputing missing continuous data in the context of a meta-analysis have received little attention in the methodological literature. In some situations it may be possible to exploit standard (although often questionable) approaches such as ‘last observation carried forward’, or, for change from baseline outcomes, to assume that no change took place, but such approaches generally require access to the raw participant data. Inflating the sample size of the available data up to the total numbers of randomized participants is not recommended as it will artificially inflate the precision of the effect estimate.
A simple way to conduct a sensitivity analysis for continuous data is to assume a fixed difference between the actual mean for the missing data and the mean assumed by the analysis. For example, after an analysis of available cases, one could consider how the results would have differed if the missing data in the intervention arm had averaged 2 units greater than the observed data in the intervention arm, and the missing data in the control arm had averaged 2 units less than the observed data in the control arm. A Bayesian approach, which automatically down-weights studies with more missing data, has been considered (White 2007).