Biotech Values

[DewDiligence]

BMTI was testing for non-inferiority, which warrants a different interpretation of the data than in a superiority study.

not stat sig… I highlighted the ITT results.

In a non-inferiority study, the ITT results are typically not the most consequential ones. This is easier to see if one bears in mind that a non-inferiority study is effectively a superiority study with a pre-agreed non-inferiority margin.

The following discussion pertains to non-inferiority trials in general as well as to BMTI’s trial in particular. Without loss of generality, we can simplify the discussion by assuming a trial is randomized 1:1 to an experimental arm and a control arm. In a regular superiority study, a pair of patients consisting of one patient from each arm with null data lowers that statistical power of the trial on an ITT basis, and hence it works to the disadvantage of the sponsor.

In a non-inferiority trial, on the other hand, a pair of patients consisting of one patient from each arm with null data raises the statistical power of the trial on an ITT basis because the null differential between the two patients in the pair falls within the pre-agreed non-inferiority margin. That null data could work to the sponsor’s advantage on an ITT basis introduces a potential bias in the conduct of a non-inferiority trial.

The above is even easier to see by carrying the argument to its logical conclusion. Consider a hypothetical non-inferiority trial in which every patient dropped out immediately after randomization—before even being dosed or treated. All patients would have null data and there would be no observed differences between the two trial arms; the pre-agreed non-inferiority margin would thus be satisfied trivially. On an ITT basis, the number of patients in the data set would not be reduced by the discontinuations, and the result of the study would be a statsig finding of non-inferiority.

To avoid the bias described above, a modified intent-to-treat (mITT) population, which excludes patients with null data, is often preferred for analysis of non-inferiority studies.