non-inferiority with interim analysis

[nsns]

Hello,

I am running an analysis for a non-inferiority design with proportions.  The current analysis is an interim analysis where I am applying Pocock boundary (significance boundary pvalue = 0.0310).

The analysis is comparing proportion difference.  The non-inferiority margin is -0.10.

The code I am using is:

proc freq data=a order=data;
tables group*Response / riskdiff(noninf margin=.10 method=fm norisks);
weight Count;
run;

Usually in a non-inferiority setting I would compare the lower bound of the confidence interval to the non-inferiority margin.  If the lower bound is greater than margin this indicates non-inferiority.  However, because this is an interim analysis and I have a pvalue boundary, I need to compare the pvalue that is generated to the pvalue significance boundary as well. 

The SAS output from the above code shows the Farrington-Manning non-inferiority pvalue and confidence interval.  My question is what happens if the pvalue boundary is not crossed - for example if the pvalue comes out to be 0.04 (compared to boundary of 0.03) then I cannot stop the trial for efficacy - i.e non-inferiority not shown.  However the lower bound of the confidence interval is greater than the non inferiority margin - which would indicate non-inferiority has been shown.  

Is there an option in SAS that would reconcile these?  If not, which would be the more correct measure to use - the NI pvalue or the lower bound CI,  to determine whether the study can stop at the interim?

Thank you.

[SteveDenham]

Consider this:  Suppose you had 10 times as many subjects, and observed exactly the same proportions in the two groups.  I suggest that the p value would be substantially less than the Pocock boundary of 0.0310, and the lower bound of the difference would be greater than the inferiority margin.  So you would stop the trial as both criteria had been attained.

So, in the end, the data at this point do not support non-inferiority, as BOTH criteria are not met.  More data might solve this problem, except that you have already spent 0.031 of your alpha  So, in the end, your test is underpowered at this point.

SteveDenham 

[nsns]

Thanks Steve.  I agree with you and actually that is what my thoughts were i.e. that both criteria would need to be met at the interim and if not, then at the final, I would need to compare the pvalue at final look to the Pocock cutoff (in this case it would be 0.0297 - very close to 0.03 which is characteristic of Pocock) and the lower bound CI to the NI margin (which intuitively would be acceptable if the pvalue is less than 0.05).

I was just concerned that I wasn't missing an option in the Proc Freq that would make some type of adjustment to the confidence interval, but from what I understand from your reply is that this is not necessary.