Often some participants are excluded from analyses of randomized trials, either because they were lost to follow-up and no outcome was obtained, or because there was some deviation from the protocol, such as receiving the wrong (or no) treatment, lack of compliance, or ineligibility. Alternatively, it may be impossible to measure certain outcomes for all participants because their availability depends on another outcome (see Section 16.2.4). As discussed in detail in Chapter 8 (Section 8.13), an estimated intervention effect may be biased if some randomized participants are excluded from the analysis. Intention-to-treat (ITT) analysis aims to include all participants randomized into a trial irrespective of what happened subsequently (Newell 1992, Lewis 1993). ITT analyses are generally preferred as they are unbiased, and also because they address a more pragmatic and clinically relevant question.
The following principles of ITT analyses are described in Chapter 8 (Section 8.13).
Keep participants in the intervention groups to which they were randomized, regardless of the intervention they actually received.
Measure outcome data on all participants.
Include all randomized participants in the analysis.
There is no clear consensus on whether all criteria should be applied (Hollis 1999). While the first is widely agreed, the second is often impossible and the third is contentious, since to include participants whose outcomes are unknown (mainly through loss to follow-up) involves imputing (‘filling-in’) the missing data (see Section 16.1.2).
An analysis in which data are analysed for every participant for whom the outcome was obtained is often described as an available case analysis. Some trial reports present analyses of the results of only those participants who completed the trial and who complied with (or received some of) their allocated intervention. Some authors incorrectly call this an ITT analysis, but it is in fact a per-protocol analysis. Furthermore, some authors analyse participants only according to the actual interventions received, irrespective of the randomized allocations (treatment-received analysis). It is generally unwise to accept study authors’ description of an analysis as ITT; such a judgement should be based on the detailed information provided.
Many (but not all) people consider that available case and ITT analyses are not appropriate when assessing unintended (adverse) effects, as it is wrong to attribute these to a treatment that somebody did not receive. As ITT analyses tend to bias the results towards no difference they may not be the most appropriate when attempting to establish equivalence or non-inferiority of a treatment.
In most situations, authors should attempt to extract from papers the data to enable at least an available case analysis. Avoidable exclusions should be ‘re-included’ if possible. In some rare situations it is possible to create a genuine ITT analysis from information presented in the text and tables of the paper, or by obtaining extra information from the author about participants who were followed up but excluded from the trial report. If this is possible without imputing study results, it should be done.
Otherwise, it may appear that an intention-to-treat analysis can be produced by using imputation. This involves making assumptions about the outcomes of participants for whom no outcome was recorded. However, many imputation analyses differ from available case analyses only in having an unwarranted inflation in apparent precision. Assessing the results of studies in the presence of more than minimal amounts of missing data is ultimately a matter of judgement, as discussed in Chapter 8 (Section 8.13). Statistical analysis cannot reliably compensate for missing data (Unnebrink 2001). No assumption is likely adequately to reflect the truth, and the impact of any assumption should be assessed by trying more than one method as a sensitivity analysis (see Chapter 9, Section 9.7).
In the next two sections we consider some ways to take account of missing observations for dichotomous or continuous outcomes. Although imputation is possible, at present a sensible decision in most cases is to include data for only those participants whose results are known, and address the potential impact of the missing data in the assessment of risk of bias (Chapter 8, Section 8.13). Where imputation is used the methods and assumptions for imputing data for drop-outs should be described in the Methods section of the protocol and review.
If individual participant data are available, then detailed sensitivity analyses can be considered. Review authors in this position are referred to the extensive literature on dealing with missing data in clinical trials (Little 2004). Participants excluded from analyses in published reports should typically be re-included when possible, as is the case when individual participant data are available (Stewart 1995). Information should be requested from the trial authors when sufficient details are not available in published reports to re-include exclude participants in analyses.