0
$\begingroup$

In the context of group sequential methods(GSM) for a clinical trial with two treatment arms, it is usually assumed that data follows normal distribution $N(\mu,\sigma^2)$. Then GSM such as Pocock's test and O'Brien and Fleming test can be applied to the problem.

My question is that if data does not follow normality assumption, what kind of test should one use ? Since in reality many data are from skewed distributions,e.g., exponential, chi-square or lognormal, violation of normality should often be seen. Then in this case what method should we use to do group sequential test? Any reference or material would be appreciated.

thank you.

$\endgroup$
1
$\begingroup$

A key thing to realize is that in the assumptions of many sequential methods, it's not an assumption that the data is normal, but rather that the statistic is normal.

This is a much more reasonable assumption. As you stated, most analysts rarely see normally distributed data. However, it is quite common that the statistic of interest approaches a normal distribution as the sample size becomes reasonably large.

As a simple example, suppose you are comparing means between two different groups. The data itself may be non-normal, but the central limit theorem states that that the estimated means will be approximately normal for large samples.

$\endgroup$
  • $\begingroup$ Good point. I understand your point. I agree that as sample size increases, the corresponding test statistic often converges in distribution to normal.\\ But as we know that when sample size is small, the CLT (central limit theorem) often does not work well for underlying skewed distributions. So is there any nonparametric method for group sequential method? I am a newbie into area of GST. But I plan to propose a nonparametric test and now do some literature searching. $\endgroup$ – lzstat Jan 26 '17 at 3:46

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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