Log in
Register|
The Outlaw Torn ★ Europe, 2012-02-09 13:04 (5244 d 05:22 ago) (edited on 2012-02-09 13:43) Posting: # 8086 Views: 6,706 |
|
|
Hi everyone, I must first admit that I am statistically challenge. Yes, I took it at university and passed, but the keg of beer I had after the final exam purged more than the contents of my bladder. Let's say we want to conduct an equivalency study, with 80% power and confidence margins of 10%. The primary outcome we want to measure, depending on what criteria is used, occurs at a rate of either 5% or 10%. My brain is telling me that if we chose a primary outcome that will occur at an expected rate of 5%, we would need a significantly higher sample size than if we chose a primary endpoint that occurs at a rate of 10%. But I'm told that this is not actually the case. I probably misunderstood, but this is how I understood what I was told. Apparently below 50% (expected rate), the variability in expected to be high (variability is apparently higher at 10% than at 5%!), and therefore, the lower the rate the lower the sample size needed. However, once you go above 50% expected rates, then the opposite happens. As you approach 100% expected rates, then the variability descreases and the sample size decreases too (so, it seems variability is highest in the 50% neighborhood and lowest in the suburbs). The bit between 50% and 100% makes perfect sense to me. But the bit below 50% does not. Like I said, I would expect a higher sample size with a primary outcome with expected rates of 5% than with an expected rate of 10%. But that's not what I'm being told. Can someone point out where I'm going cuckoo? Thank you. |
|
d_labes ★★★ Berlin, Germany, 2012-02-10 10:28 (5243 d 07:58 ago) @ The Outlaw Torn Posting: # 8093 Views: 5,890 |
|
|
Dear Outlaw Torn, ❝ I must first admit that I am statistically challenge. Welcome to the club. Search the forum for 'statistical amateur', 'bloody amateur' or 'bloody raw recruit in statistics'  .❝ Let's say we want to conduct an equivalency study, with 80% power and confidence margins of 10%. The primary outcome we want to measure, depending on what criteria is used, occurs at a rate of either 5% or 10%. My brain is telling me that if we chose a primary outcome that will occur at an expected rate of 5%, we would need a significantly higher sample size than if we chose a primary endpoint that occurs at a rate of 10%. This is a question that is somehow beyond the horizon of us doing mainly BE studies all of our working time. Without really understanding what I did, I fired up PASS2008 [1]. Choosing the module Equivalence of proportion/Two independent proportions/Specify using differences I got: PR N (per group)I set target power to 80%, alpha=0.05, equivalence margin (for the risk difference PT-PR) +10% and 'true' difference = zero. The power/sample size was calculated based on the simplest test in this respect, the unpooled Z-test. If the sample size for the 5% rate is significantly higher then for the 10% rate is left to you  .❝ But I'm told that this is not actually the case ... So what? Show the results of my sample size estimation to your statistical consultant and ask him to comment. ❝ The bit between 50% and 100% makes perfect sense to me. But the bit below 50% does not. The binomial is 'symmetric' about 50%. If this makes sense for you or not  .Look at the sample size: PR N (per group)Again with 'true' risk difference = zero. BTW: The binomial is always good for a surprise. Have a look onto the sample size if 'true' risk difference is set to 0.01 instead of zero: PR NIn case of the 5% outcome the sample size decreases! Somehow counter intuitive for me. But as said above: Didn't really understood what I'm doing here .[1] Hintze, J. (2008) PASS2008, NCSS LLC. Kaysville, Utah — Regards, Detlew |
|
The Outlaw Torn ★ Europe, 2012-02-10 12:59 (5243 d 05:28 ago) @ d_labes Posting: # 8095 Views: 5,755 |
|
|
Dear d_labes, Thanks for the feedback. Again, everything made sense up until you brought in that darn binomial—then the earth's polarity changed. Somedays I think I never should have left that pub after the exam.  |
|
d_labes ★★★ Berlin, Germany, 2012-02-10 14:10 (5243 d 04:16 ago) @ The Outlaw Torn Posting: # 8098 Views: 5,689 |
|
|
Dear Outlaw, ❝ ... Somedays I think I never should have left that pub after the exam. I strongly suggest you to take the opportunity to correct your former decision before 31-Dec-2012 .— Regards, Detlew |
|
The Outlaw Torn ★ Europe, 2012-02-10 15:18 (5243 d 03:09 ago) @ d_labes Posting: # 8099 Views: 5,795 |
|
|
Ah, the Mayans! Took me a while to get it.  I need a beer bad, as you can tell.Which method is right for a clinical trial to test equivalence of a generic to a reference product—proportions or binomial? |
|
d_labes ★★★ Berlin, Germany, 2012-02-10 16:18 (5243 d 02:08 ago) @ The Outlaw Torn Posting: # 8100 Views: 5,775 |
|
|
Dear Outlaw ❝ Which method is right for a clinical trial to test equivalence of a generic to a reference product—proportions or binomial?  Proportions are distributed according to binomial distribution (or if the events you are observing are very rare according to the Poisson distribution).So if you talk about comparison of proportions you talk about comparisons of binomial outcomes. Hope this doesn't add to your beer thirst .Have a nice weekend. — Regards, Detlew |
Ing. Helmut Schütz