I'm afraid I can't get on board with TechGuru's discussions of post-hoc power. I have never seen it calculated when a result is significant, and there are problems in using it even when p>.05.
When we began CD-12, best guess at the power calculation was that we had 90% odds of detecting a 30% death rate reduction. You CAN NOT specify power without specifying the alternative you care about.
Yes, that depended on a certain death rate in the control group. Post hoc you can recalculate, based on actual total deaths, whether the assumption converting sample size to total deaths was wrong. That's what the DSMBs do in deciding if sample size should be increased to "maintain power."
KEY POINT: this calculation is still vs. the 30% death reduction target, and it has nothing to do with the observed difference between treatment vs. control death rates.
As TechGuru is calculating it, I'm pretty sure any outcome with P=.05 will calculate to a 50% post-hoc power.
There is a concept called the fragility index that does apply.
When we began CD-12, best guess at the power calculation was that we had 90% odds of detecting a 30% death rate reduction. You CAN NOT specify power without specifying the alternative you care about.
Yes, that depended on a certain death rate in the control group. Post hoc you can recalculate, based on actual total deaths, whether the assumption converting sample size to total deaths was wrong. That's what the DSMBs do in deciding if sample size should be increased to "maintain power."
KEY POINT: this calculation is still vs. the 30% death reduction target, and it has nothing to do with the observed difference between treatment vs. control death rates.
As TechGuru is calculating it, I'm pretty sure any outcome with P=.05 will calculate to a 50% post-hoc power.
There is a concept called the fragility index that does apply.
5
