## Power in Diletti's sample size table [R for BE/BA]

Dear all,

I performed some comparisons of the code based on Jones/Kenward implemented in R, Diletti's table and results obtained from StudySize (v2.0.1). Quite interesting...

All comparisons were done for CV=20% with ratios of 0.85-1.20. Dilleti reported sample sizes to obtain >80% power, calculated odd sample sizes were reported rounded up to the next even number (underlined):
 +======+=====+===========+===========+ |  GMR |  n  |  R-code   | StudySize | +------+-----+-----------+-----------+ | 0.85 | 134 | 0.8014178 |  0.80167  | | 0.90 |  38 | 0.8140704 |  0.81536  | | 0.95 |  20 | 0.8300156 |  0.83451  | | 1.00 |  16 | 0.8214263 |  0.83305  | | 1.05 |  18 | 0.7950343 |  0.79996  | | 1.10 |  32 | 0.8084890 |  0.80992  | | 1.15 |  72 | 0.8035456 |  0.80411  | | 1.20 | 294 | 0.8017617 |  0.80182  | +======+=====+===========+===========+

Power with sample sizes given by Diletti et al. at a GMR of 1.05 were below 80%, calculated both with R and StudySize...

A Monte Carlo Simulation (1000000 runs) for GMR=1.05 and 18 subjects in StudySize resulted in:
Power 0.8006 (95% coverage probability: 0.7998-0.8013).

Differences may be due to the implementation of the algorithm to obtain the value of the noncentral t-distribution by numeric integration...

Maybe somebody of you has access to SAS or software specialized in power analysis (e.g., PASS or nQuery Advisor) and would like to check these results?

Cheers,
Helmut Schütz

The quality of responses received is directly proportional to the quality of the question asked. ☼
Science Quotes