On Fri, 25 Sep 2020 03:08:58 +0000 (UTC), David Duffy
<
[email protected]> wrote:
Cosine <[email protected]> wrote:
Hi:
What do we prove by showing the statistical significance of the
results of a clinical trial?
There is a modern trend /away from/ "statistical significance."
Perhaps it is clearer to speak of the (often, 95%) Confidence
Interval, as a clearer statement of the same information. See
David Duffy's example, below.
Herman Rubin, a wise gentleman who used to post here,
regularly returned to "decision theory" as the underlying
guide. I liked that reminder. If you want to, decisions can
explicitly include costs and benefits, and Bayesian methods
sometimes make use of "informative" prior information (instead
of taking a technically-uninformative prior in the math).
Say, we have a new drug, a new device, or
a new procedure, and then we design a clinical trial to show that there
is statistical significance between the treatment results of the group
using the new stuff and those of the controlled group. If this trial
success, we claim that the new stuff is effective.
But how reliable is this claim? Say, we have no doubt that if we cut
off the head of a person, this person shall die. But even if we have new
stuff passing the trial mentioned above, in practice, we still find that
using that new stuff does not save everyone.
So, what did we prove by conducting that kind of trial?
When we look at the CI, we should also apply the notion of
"effect size." We might decide not to use a "better" method when
it is expensive and when its improvement, as shown in trials
with huge Ns, is very small.
You need to read some textbooks. Consider my head chopping off
experiment:
N Prop_died 95%_CI
5 100% 57%-100%
10 100% 72%-100%
20 100% 83%-100%
When should the Data Monitoring Committee suggest we stop the trial?
One might reach these quandries when ignoring costs and
benefits from the start.
--
Rich Ulrich
--- SoupGate-Win32 v1.05
* Origin: fsxNet Usenet Gateway (21:1/5)