Randomization is the only scientific method that can balance the representation of known and unknown variables. Because the exclusion of non-compliant patients from the analysis is likely to influence the statistical power and the balance between the study groups, the intention-to-treat concept was introduced.
Intention to treat analysis depends on the rule that “All randomized, as analyzed” where all randomized patients are included in the final analysis irrespective of any noncompliance or withdrawal from the trial.
Noncompliance to treatment
The efficacy of the intervention is more likely to be less in non-compliant patients as compared to compliant patients. Therefore, for explanatory analysis, aiming at investigating the underlying biological effects of an intervention, we should not include non-compliant patients or drop outs because it is not likely that the treatment has effect in this population even if their data were accessible. On the contrary, the ITT approach is used as a pragmatic approach to estimate the efficacy of the drug in compliant and non-compliant patients. This pragmatic approach is more relevant to clinical practice where many patients do not strictly comply to the treatment.
The problem of ITT is that end-point data of non-compliant or lost patients might not exist. Therefore, the inclusion of their data in the final analysis will be problematic. Multiple approaches have been proposed to overcome this issue:
- LOCF analysis (last observation carried forward)
In this approach, investigators use the last observation data as an end-point for lost patients. However, the problem with this scenario appear when the underlying disease has a progressing nature, meaning that the disease deteriorates over time. For example: in an RCT about neuroprotection against Parkinson’s disease, when the investigators use the last observation value, it is likely that lost patients will have values that indicate less disease progression than the actual end-point. Whatever the treatment they receive (active intervention or placebo), the earlier values are more likely to be better which might influence the effect estimates.
- Multiple imputation
In this approach, imputations are performed through regression models and random errors are added to the expected values through a random number generator.
- Analysis of the worst case scenario
If the outcome is dichotomous (e.g. Mortality), then we can assume the worst event for drop outs of the experimental group and the best event for the drop outs of the control group. If the outcome is continuous, we can assign the best baseline value and the worst endpoint value to the drop outs. Therefore, if the experimental group was found superior to the control group, we will be sure of this superiority. While if the experimental group was not superior to the control group, we will not be sure that the intervention was not effective because the type of analysis might underestimate the magnitude of treatment effect. It is not reasonable to underestimates drug efficacy due to non-compliance of patients (poor compliance ≠ less efficacy).
Because there are multiple methods to estimate missing data, it is recommended that the investigators (1) try to obtain the data of drop outs from other sources (e.g. death registry); (2) try to impute the missing data using multiple approaches; and (3) perform a sensitivity analysis analyzing complete data only in one scenario then in the ITT scenarios. This will estimate the drug efficacy pragmatically and explanatory.