topic Re: Proc Power: How to specify the type I error level when testing for equivalence in Statistical Procedures

Proc Power: How to specify the type I error level when testing for equivalence

AVoelp — Fri, 24 May 2013 14:40:38 GMT

I am planning the sample size for an equivalence trial with two parallel groups and a metric clinical outcome (comparison of mean values). The type I error level of the (two-sided) equivalence problem will be alpha=0.05, corresponding to the inclusion of the two-sided 95% conficence interval for the point estimate between the lower and the upper bound of the equivalence range. The null hypothesis will be tested by performing two one-sided tests against the lower and the upper bound, each with a one-sided type I error level of alpha/2 = 0.025.

When I calcuilate the sample size using Proc Power as in

proc power;

twosamplemeans test = EQUIV_diff

meandiff = 0

stddev = 25

npergroup = .

lower = -10

upper = 10

alpha = 0.025 0.05

power = 0.8;

run;

... do the values that I specify for alpha correspond to the two-sided case (i.e., to the procedure as a whole), or to the one-sided error level of each of the two one-sided tests? Which value for alpha will be appropriate for the test problem described above?

Re: Proc Power: How to specify the type I error level when testing for equivalence

SteveDenham — Thu, 30 May 2013 19:03:13 GMT

Good question. I have always treated them as corresponding to the two-sided case. If this is incorrect, I hope someone can let us know.

Steve Denham

Re: Proc Power: How to specify the type I error level when testing for equivalence

AVoelp — Fri, 31 May 2013 07:21:35 GMT

Steve, you're on the safe side

Since there have been no responses for days (from which I have concluded that the topic is either not of interes, self-evident, or nobody really knows), and as I do not have access to SAS myself, I have cintacted their support, and they have oinformed me that the two-sided case is the program default. You can override this by specifying 'sides=1'.

Andreas