topic Re: Sample size for one-proportion binomial noninferiority test in Statistical Procedures

Sample size for one-proportion binomial noninferiority test

newtriks — Fri, 16 Aug 2024 19:10:50 GMT

Running a power analysis on 9.4 to determine sample size. Comparing a new testing method to a "gold standard" test which is correct 100% of the time on a binary, yes/no outcome. Will accept the new method as "noninferior" if it matches the gold standard 90% of the time. One-sided Alpha = .05 with 90% power.

I'm having difficulty with the coding. I went with what I thought the SAS manual recommended for the above parameters, but it won't give me an N ("ERROR: NTOTAL is not available for TEST=EXACT").

PROC POWER;

ONESAMPLEFREQ

TEST = EXACT

METHOD=EXACT

NULLPROPORTION=.9

PROPORTION = .8 TO .95 BY .01

SIDES = U

POWER = .9

ALPHA = .05

NTOTAL = .;

PLOT X = EFFECT'

RUN;

Any help would be appreciated. Thanks.

Re: Sample size for one-proportion binomial noninferiority test

newtriks — Fri, 16 Aug 2024 19:20:02 GMT

There's a semicolon after PLOT X = EFFECT; - that's not what's causing the issues.

Re: Sample size for one-proportion binomial noninferiority test

ballardw — Fri, 16 Aug 2024 22:17:20 GMT

When SAS tells you it won't do something, such a NTOTALS for TEST=Exact, then it won't.

But sometimes you can approach the problem from the other side. Provide a range of values for NTOTALS and see what the Power is. Refine the lists for those cases of interest. As an example of a starting point.

PROC POWER;
ONESAMPLEFREQ 
TEST = EXACT
METHOD=EXACT
NULLPROPORTION=.9
PROPORTION = .8 TO .95 BY .01
SIDES = U
POWER = .
ALPHA = .05
NTOTAL =  100 to 1000 by 100;
RUN;

You can see that for proportions under .90 that 1000 is too small.

You also get notes

WARNING: The effect is in the opposite direction from the alternative hypothesis in one of the
         scenarios in the output table.

actually more than one of the scenarios. Which means you may to reconsider if all of your combinations of NULLPROPORTION and PROPORTION actually make sense.

Play with values of NTOTALS and see how close you get to the power of .9 you can get. And which scenarios NEVER get there.

Re: Sample size for one-proportion binomial noninferiority test

newtriks — Sat, 17 Aug 2024 13:06:53 GMT

Thanks!

Re: Sample size for one-proportion binomial noninferiority test

Ksharp — Sun, 18 Aug 2024 02:02:20 GMT

Since you are doing one-way noninferiority test,

You should firstly specify a decent MARGIN value in PROC POWER,

And adjust parameter proportion= ,margin= ,ntotal= to get rid of these WARNING message.

proc power;
onesamplefreq test=exact
sides = U
proportion = .82 TO .9 BY .01
nullproportion = 0.9
margin = -0.1
ALPHA = .05
ntotal =  100 to 1000 by 100
power = .;
run;

Re: Sample size for one-proportion binomial noninferiority test

newtriks — Tue, 20 Aug 2024 15:17:37 GMT

Thanks. I tried to accept this as a solution.

Re: Sample size for one-proportion binomial noninferiority test

newtriks — Tue, 20 Aug 2024 15:20:07 GMT

Thanks. Tried to accept as solution.