Sample size calculation for proportion repeated measures

[Seth]

Hi...I'm trying to figure out which procedure to use to calculate sample size for a difference in difference binary repeated measures. It looks like proc GLMPOWER would work great, as the examples shown match my situation quite well. The only exception is i have proprotions in my two time periods (pre and post). Again, i am trying to determine the difference for say the control going from 50% to 55% (pre to post) versus the test period going from 50% to 57% (pre to post). I know the treatment*time term will give us this answer, but i am trying to find the correct procedure to calculate sample size in this prospective study.

This is what the hypothetical summary data looks like:

cntl 0.50 0.55

test 0.50 0.57

Thanks

[Damien_Mather]

do you mean that you want to know the sample sizes required with enough sensitivity to pick up at least a 5% and 7% increase in binary responses respectively?

[Seth]

no, this is a difference in difference, where i want to look at the time*treatment interaction term...this term parses out the effect due to just treatment, controllng for the pre post time trend. A mixed model, or a two way ANOVA with repeated value is the way to conduct analysis, but i am looking for samples size calculation for this prospective study. GLMPower seems to be the new way to handle sample calculations, but it doens mention if it applies to binary data or not.

here is SAS code that i used. Again, my question is does PROC GLMPoser apply to binary data, and if not, how bad of an error would it be to use GLMPower to binary data?

data TrainingSummaryData;

input Training $ Pre Post;

datalines;

test 0.70 0.75

cntl 0.70 0.7875

;

run;

ods graphics on;

proc glmpower data=TrainingSummaryData;

class Training;

model Pre Post = Training;

repeated Time contrast;

power

mtest = hlt

alpha = 0.05

power = .8

ntotal = .

stddev = 0.25 0.5 1.0

 matrix ("Corr6") = (0.6)

/* corrmat = "Corr";*/

corrs = "Corr6" ;

run;

[sld]

If you anticipate doing data analysis with the GLIMMIX procedure (or, I suppose, even if you do not), you can specify an exemplary data set of interest and use the HOLD option on the PARMS statement to compute power. Walt Stroup illustrates the process here

http://support.sas.com/resources/papers/proceedings16/11663-2016.pdf

The method is also covered in his text (Generalized Linear Mixed Models, CRC Press, 2013, Ch 16).

I think GLMPOWER is appropriate for response variables that follow normal distributions, not for proportion or binomial responses. The POWER procedure has capabilities for non-normal responses, but minimal abillity to deal with mixed models.

[Damien_Mather]

Thanks for the reference to the paper, what a gem that is. I'll be using that method for a lot of my designs in future.