Demystifying intention-to-treat analyses
In RCTs, participants are randomly allocated to compared groups; thus, in the simple case of a trial with only two groups (two trial arms), some participants are allocated by a chance (statistical) mechanism to receive an intervention of interest and some participants are allocated to a comparator group. Ideally, in a well executed trial, all participants allocated to a group remain in that group for the entire planned duration of the trial, receive the intended intervention and at the planned dosage/intensity and frequency and are fully compliant with the intervention and all the trial procedures; ideally, the same happens for the control group. In these ideal conditions, given that the compared groups determined by the random allocation are assumed comparable with regard to known and unknown confounders (covariates), it is possible to rule out other explanations for the difference between the effects in the intervention and the control groups, and consider the intervention as the plausible cause of the observed effect. In these ideal circumstances, randomization justifies and assures the conditions for the use of statistical procedures (such as significance testing and computation of P values or the estimation of confidence intervals), and a statistical analysis may provide an estimate of the effect of receiving the intervention in conditions of 100% compliance with the treatment and complete follow-up.
In reality, clinical trials are far from these ideal, perfect conditions: some participants leave the assigned group for various reasons, follow-up is incomplete, some participants receive an incomplete intervention (e.g. lower dosage or shorter duration than planned), other participants switch to the control intervention, participants are compliant with the interventions to variable degrees (e.g. some participants are fully compliant, some completely not compliant and some participants somewhat compliant) and so forth. In short, the group of participants receiving the intervention of interest and the comparator group may not reflect the groups determined by randomization; furthermore, the interventions may not be those planned. Moreover, if statistical procedures are performed on groups other than those determined by the randomization, the benefits of randomization that have been indicated above are lost.
No less than seven different statistical approaches are described for the analysis of RCTs, including, for example, analysis of the treatment effect for the treated, the analysis of treatment effect in the untreated and the ITT analysis.1 The ITT is an analysis of an RCT, in which all randomized participants are included and analyzed in the groups to which they are randomly assigned – regardless of whether they remain in those groups for the duration of the trial, receive the intervention as planned or whether they are compliant with the intervention.2-5 In this way, the ITT preserves the benefits of random assignment for causal inference as it is not estimating the effect of actually receiving the intervention of interest; rather, the ITT is estimating the effect of offering the intervention. The ITT provides “an (unbiased) answer” only to one question, specifically, “What are the expected outcomes for a typical patient instructed, in the context of the trial, to take the treatment to which he was assigned?”6(p.