What is an intention to treat analysis?
Intention to treat analyses are prevalent in randomised controlled clinical trials. Hollis and Campbell (1999) present an overview of their use and report around 50% of studies examined used intention to treat (ITT). The main points that they raise are detailed below.
The object of ITT is to use data from all subjects originally allocated to the groups at the outset of the trial (ie who were intended to be treated in the trial) irrespective of whether they dropped out or staisfied entry criteria.
Failure to include all patients can underestimate the effectiveness of a treatment. For example suppose some patients are given a memory aid which includes a reminder to 'take your medication'. The patients are not reminded of this due to a failure to use the memory aid correctly and then withdraw from the study for health reasons. This represents informative dropout in that the reason for the dropout is related to the effectiveness of thge memory aid being assessed. If these patients are excluded the usefulness of the memoru aid may be overestimated.
Care should be taken to minimise missing responses. If missing responses do occur, however, the most common ways of handling these are using complete cases (49% recorded by Hollis and Campbell (1999) and, if imputation is used, carrying forward the last recorded response (LOCF=last observation carried forward, 7 studies reported by Hollis and Campbell). The latter method has the advantage of including all the patients (so does not violate the intention to treat) and can be seen as a conservative assessment of change in that improvement may be expected to get more pronounced over time. For example suppose we are interested in improvement over time in recall using a memory aid. It might be expected that with practice people might recall more events using the memory aid so that using earlier responses denoting the number of items remembered may underestimate the number that would have been remembered later with more practice using the memory aid in those who had missing data. Other methods used for imputation in the survey of Hollis and Campbell in addition to LOCF were allocating the worse outcome (to underestimate the effectiveness of the treatment, used in 4 studies), working out rates including subjects with missing outcomes in the denominator only (3 studies) and the group average (1 study).
LOCF is the most pragamatic approach in clinical trials because these usually do not collect sufficient data to allow good estimation by more complex imputation methods. In addition, assumptions made by more complex imputation proecedures are difficult to verify in most clinical trials. Extreme case analysis (for example, all patients lost to the group that fared better are assigned a poor outcome; all lost to the group that fared worse are assigned a good outcome) has also been recommended (Sackett et al (1997)) but this is unlikely to yield a conclusive answer in practice (Meyer K, Windeler J, 19th International Society for Clinical Biostatistics meeting, Dundee 1998). This approach motivates uing sensitivity analyses to comapre various imputation scenarios in assessing the effectiveness of a treatment.
Hollis and Campbell use 10% as a cut-off for assessing the proportion of responses which had missing data in the study.
If the reason for missing data can be shown not to be connected to the response of interest e.g. non-compliance due to issues which were not connected to the response including factors which occur unexpectedly which are beyond the control of the study e.g. someone regaining their memory or not using the memory aid correctly then these may be considered for dropping from the analysis. Such an analysis, however, would not be regarded as an intention to treat since not all subjects were used. It may be possible
Sackett DL, Richardson WS, Rosenberg WS and Haynes RB (1997). Evidence-based medicine. Churchill Livingstone:New York.