"I think it is quite correct to say that if they could have a bigger trial, they may find a benefit for the new drug. Where's the catch ?" I was puzzled.
"You are caught by doing correct mathematics," my friend twisted his lips, "Look, the study that we are talking about has over 1600 subjects. What benefit could it not detect ?"
"Why, it depends on both the magnitude of the benefit and the prevalence of the problem." I still remembered the basic principles of sample size estimation.
"Of course, of course. As I said, you are obsessed with numbers and details," L was not moved, "Say, if the outcome would develop in half of the untreated subjects, what benefit of a new treatment could be reliably detected by a clinical trial of 1600 patients ?"
"You are pulling my legs. I do not bring along with me the software." I complained.
"Don't worry. Let me show you this," and then he gave me a tiny sheet of note, on which was full of his scribbles.
Here it goes:
- assume 80% power and a P value of 0.05 being statistically significant
- two-arm study; 800 subjects each side
- if the risk of developing an event in the control arm is 50%, the study has sufficient power to discern it from a treatment arm with an event rate of 43%
- in other words, it can detect an absolute risk reduction of 7%, or relative risk reduction of around 14%
- if the treatment has the benefit of a relative risk reduction by 50%, a sample size of 1600 would have sufficient power to detect the difference when the risk of event is 6% (i.e. the treatment reduce the prevalence to 3%)
"The point is, if you need an astronomical sample size to detect a difference, either the risk of event is very low, or the benefit is really marginal, or, more likely, both," my friend sighed, "And the concern is not only relevant to this fortunately harmless medicine - it applies to all mega-trials !"
1 comment:
And that's why I have been preaching to the students that they should ignore all mega-trials!
Post a Comment