Non-inferiority Analysis

[Sara Hansen]

Hi,
I want to make a non-inferiority analysis, investigating if treatment A is not worse than treatment B. In my trial I only test treatment A, the results for treatment B is from another published study. Treatmen B's mean = 55.
I have used stata's ssi module to make my power calculation, and now want to make sure to make the analysis correctly. The chosen non-inferiority margin is 48, i.e. 7 points lower than Treatment B's 55.

Since it is a non-inferiority study, my hypothesis is tested at a 2.5% significance level by the upper 97.5% CI (single sided), which is the upper limit of the two sided 95% CI. - and 95% CI is the default for stata's ttest.

ttest treatmentA == 55

If the lower bar of the CI is above 48, I can conclude that treatment A is non-inferior to treatment B, right?
- How could I from the same CI conclude, if treatment A was not just non-inferior, but actually superior?



Another question: is it correctly understood, that the p-value isn't really used in non-inferiority testing; only the confidence intervals? - And if so, should one also report the p-value or simply omit it?


Thank you for your time.

[Carlo Lazzaro]

Sara:
Joseph Coveney might have posted about this issue. Skim though his replies.

[Joseph Coveney]

If the lower bar of the CI is above 48, I can conclude that treatment A is non-inferior to treatment B, right?

If you're treating mean value out of the literature as a known fixed parameter, that is, if you're considering the variability in the estimate as ignorable, then yes.

How could I from the same CI conclude, if treatment A was not just non-inferior, but actually superior?

Your lower confidence bound must exceed 55, not just 48, for that claim.

is it correctly understood, that the p-value isn't really used in non-inferiority testing; only the confidence intervals? - And if so, should one also report the p-value or simply omit it?

In this context, aren't they flip sides of the same coin? (Google TOST two one sided t-test for a classic example.) As for reporting I recommend following your audience's expectations of what's to be reported.