Sample size calculation for repeated measure

[ph6]

I need to calculate the sample size for a study of biomarkers which indicate the progress of a disease. The samples will be collected at baseline, 12, 24,36 and 48 months and will be tested against 15 CSF biomarkers. A pilot study has been done with 15 patients which I would like to use as prior information. I believe that the biomarkers are independent but there also could be some correlation. The study question is to see if the continuous measurements of each biomarker changes over time. The biomarkers values are not normally distributed. I have tried Gpower software using non parametric Anova for repeated measure. My first question is whether I need to calculate the sample size for each biomarker individually using Bonferroni correction 0.05/15 and then choose the largest sample size, or to calculate the sample size for all biomarkers together. Another question, for the gpower I would need, no sphericity correction, epsilon, correlation between time points and eta square, I tried to calculate them using proc glm, as

proc glm data=data;
class timepoint;
model biom= ...
repeated timepoint/ Printe;
run;

But didn't get what I want. Do you know how can I obtain those values and shall I assume data for each biomarker separately?

[jiltao · Accepted solution]

You might consider using PROC GLMPOWER for the power and sample size computations for ANOVA models. Here is an example --

https://go.documentation.sas.com/doc/en/pgmsascdc/v_017/statug/statug_glmpower_examples03.htm

Thanks,

Jill

[ph6]

Thanks a lot for this link. Indeed very helpful. I have a pilot study from which I would like to estimate the error standard deviation and correlation matrix. I use 

```

proc corr data = ccl2_t;
var _0 _12 _24 _36 _48;
run;

```

to obtain the correlations but I get some negative values which do not suit for the proc glmpower. Do you know how that can be fixed? How can I simply enter the correlation matrix?

For the error standard deviation I simply find the SD of the mean differences. Is that a right approach?

[jiltao]

You can enter negative correlation values in PROC GLMPOWER. The documentation below shows various ways of entering values for the MATRIX()= option --

https://go.documentation.sas.com/doc/en/pgmsascdc/v_017/statug/statug_glmpower_syntax11.htm#statug.g...

The STDDEV= is the error standard deviation, or root MSE. You might considering fitting a PROC GLM model to a reasonable data and use the root MSE from the output.

Thanks,

Jill

[SteveDenham]

If you have good measures of the residual error and the correlations, consider using the method developed by Walt Stroup, and outlined in Chapter 12 of SAS for Mixed Models, 2nd ed.  The book provides an analytic method using PROC MIXED and then works through a simulation method.

SteveDenham

[kjmathes03]

Hello Steve,

Any suggestions if I don't have access to SAS for Mixed Models, 2nd ed.  ?