The Cochrane Collaboration open learning material  The Cochrane Collaboration web site
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 A1 A2
Module contents:
Further issues in meta-analysis
Learning objectives
Sensitivity analysis
Other types of data
Intention to treat issues
Indirect comparisons
Next module


Intention-to-treat: analyze participants in the groups to which they were randomized

Intention to treat issues

Intention-to-treat (ITT) analyses are widely recommended as the preferred approach to the analysis of most clinical trials. Systematic reviewers often wish to practice this recommendation and plan to conduct meta-analyses according to the ITT principle. But what does this mean, and how might it be achieved?

The ITT principle

The basic intention-to-treat principle is that participants in trials should be analysed in the groups to which they were randomized, regardless of whether they received or adhered to the allocated intervention. Two issues are involved here. The first issue is that participants who strayed from the protocol (for example by not adhering to the prescribed intervention, or by being withdrawn from active treatment) should still be kept in the analysis. An extreme variation of this is participants who receive the treatment from the group they were not allocated to, who should be kept in their original group for the analysis. This issue causes no problems provided that, as a systematic reviewer, you can extract the appropriate data from trial reports.

The rationale for this approach is that, in the first instance, we want to estimate the effects of allocating an intervention in practice, not the effects in the subgroup of participants who adhere to it.

The second issue in ITT analyses is the problem of loss to follow-up. People are lost from clinical trials for many reasons. They may die, or move away; they may withdraw themselves or be withdrawn by their clinician, perhaps due to adverse effects of the intervention being studied.

If participants are lost to follow-up then the outcome may not be measured on them. But the strict ITT principle suggests that they should still be included in the analysis. There is an obvious problem - we often do not have the data that we need for these participants. In order to include such participants in an analysis, we must either find out whether outcome data are available for them by contacting the trialists, or we must 'impute' (i.e. make up) their outcomes. This involves making assumptions about outcomes in the 'lost' participants.

There are many 'formal' approaches to imputing missing outcomes in clinical trials. A review of these is beyond the scope of this course. We shall look at one particular situation that arises commonly and consider some alternative approaches that might be compared in a sensitivity analysis.

Consider the following trial of a Larium-Qinghaosu combination versus Qinghaosu alone for treating malaria that was included in a Cochrane review. Although 20 were randomized to the former and 18 to the latter, results were available only for the 34 people that did not drop out. These were the findings regarding the presence of parasitic infection after four weeks:

  Clear Not clear Total
Larium+Qinghaosu 17 0 17
Qinghaosu 10 7 17
Several methods of filling in missing outcomes can be compared in a sensitivity analysis

There were two other similar trials, also with missing data. In order to perform an ITT analysis, these data needed to be imputed. Four particular strategies for doing this are

  1. (Assume the best) assume everybody missing was clear of infection
  2. (Assume the worst) assume everybody missing was not clear of infection
  3. (Best-case scenario for combination treatment) assume everybody missing on the combination treatment was clear and everybody missing on Qinghaosu was not clear
  4. (Best-case scenario for single treatment, or worst-case scenario for combination treatment) assume everybody missing on the combination treatment was not clear and everybody missing on Qinghaosu was clear

Strategies 1 and 2 are attempts to fill in the missing data in a realistic manner. One of these may be an obvious contender for your situation. For example, in clinical trials of smoking-cessation treatments it may be reasonable to assume that dropouts continue to smoke.

Strategies 3 and 4 put bounds on the possible results of the trial had all participants been observed. Fortunately in the malaria review the results were not sensitive to any of these strategies. When dropout rates are higher the results may not be robust and caution may be needed in interpreting the findings.

If your outcome is a continuous measure, imputation of missing data for ITT purposes is more difficult, as there are more than two different possibilities for each participant.

Because imputation of missing data in order to perform a full ITT analysis is controversial, it may be best to present only the results for available participants. If you do this, you should also consider the possible effects of the missing participants, either through sensitivity analyses as described here or by discussing the implications in the Discussion of your review. An alternative approach may be to only analyse the data available, but to consider drop out rate as a marker of trial quality. Whichever approach you use, ensure that it is described in the methods section of the review and that the numbers of participants with missing data are described in the results section and the characteristics of included studies table.

© The Cochrane Collaboration 2002   Next: Indirect comparisons