topic Re: Power Estimate with Standard Deviation Percentage Difference in Statistical Procedures

Power Estimate with Standard Deviation Percentage Difference

mekono — Mon, 02 Aug 2021 22:43:42 GMT

I want to estimate the power associated with various treatment effects between two groups. In group 1 there are 145 people and group 2 has 225 people. Additionally, I expect the average score of group 1 to be 31% and group 2 41%. Based on previous studies, we expect a 5% standard deviation in each cohort. How can I incorporate 5% sd into my power estimate?

Originally, I estimated power this way:

proc power;
twosamplemeans test=diff
groupmeans = 0.31 | .41
stddev = 0.05
groupns= (145 225)
power = .;
run;

Is this correct or should I take another approach?

Re: Power Estimate with Standard Deviation Percentage Difference

sbxkoenk — Mon, 02 Aug 2021 22:59:46 GMT

Hello @mekono ,

I have never really used PROC POWER, but what SAS-version are you using?

SAS 9.4 M7 (SAS/STAT 15.2) has a Power and Sample Size Application. Also available in earlier versions of SAS 9.4, I believe.

The functionality of the Power and Sample Size application has been replaced by tasks in SAS Studio.

So, if you use SAS VIYA, there are tasks for this.

Maybe, by using an application or tasks, your doubts (about the correct wording = coding) will vanish??

Good luck,

Koen

Re: Power Estimate with Standard Deviation Percentage Difference

mekono — Mon, 02 Aug 2021 23:08:17 GMT

I have SAS 9.4
I’ve been using proc power to estimate power/sample size in the past.

Re: Power Estimate with Standard Deviation Percentage Difference

jiltao — Tue, 03 Aug 2021 13:13:21 GMT

Assuming 5% is in the unit of average score (not 5% of the mean, for example), your code looks reasonable to me.

Re: Power Estimate with Standard Deviation Percentage Difference

SteveDenham — Tue, 03 Aug 2021 16:40:31 GMT

If you are concerned about the power estimate being >0.999, consider that your mean estimates differ by exactly 2 standard deviations. However they differ by 18.7 * pooled standard error. If you plug all of this into formulas in a DATA step, where you use the null difference (=0) as the true difference for calculating the noncentrality parameter for the t distribution, you'll see this is correct.

SteveDenham

PS. The noncentral parameter for a null difference of zero is, itself, zero.