I have a sample of 1000 observations, divided into 2 groups: treatment and control (500 in each group). Each observation is dichotomous: 1 if an event occurred (bad event), and 0 if no event occurred.

I wish to test the non-inferiority hypothesis that P(treatment)-P(control)=0.01 vs. P(treatment)-P(control)<0.01, meaning, to show that the proportion of these bad events in the treatment group is not larger by more than 1% from the control group.

This is a one-sided test, and therefore the significance level for the test is 2.5% and not 5% (FDA requirement).

I wish to replace the test with a confidence interval. The idea is to calculate the 95% confidence interval to the proportions difference. If the upper limit of the CI is lower than 0.01, I declare non-inferiority, and otherwise, I fail to declare non-inferiority.

I wanted to ask you, if calculating a 95% CI for proportions difference, and using it's upper limit, is as identical to doing a one-sided test with a significance level of 2.5%? Is it just as conservative as the test?

I saw some references according to which this procedure is done, but I ain't sure.

Thank you in advance.


1 Answer 1


To make things simpler I don't consider any correction here.

The confidence interval at level $1-\alpha$ corresponding to one-sided test is

$ (-\infty ; (pT - pC) + z_{1-\alpha}s )$

where $s^2 = p(1-p)(1/n_1 + 1/n_2)$; $p$ being estimated as the pooled proportion.

The same CI($1-\alpha$) corresponding to the 2-sided test is

$( (pT - pC) - z_{1-\alpha/2} s ; (pT - pC) + z_{1-\alpha/2}s ).$

Thus to get the upper bound of the one sided CI($1-\alpha$) one should use a two sided CI($1 - 2\alpha$)

For a one-sided CI(97.5%) [i.e. $\alpha$=2.5%] one should use a two sided CI(95%).

Therefore, I guess the answer to your question is yes :-)


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.