> it seems like even simple techniques could extract a lot of signal from the noise, which is just too difficult with the tiny number of low sample, high variance RCT studies we currently use.
On the other hand, RCTs gather relatively well-conditioned data. And we have enough study power across them that we end up with a lot of statistically significant, real findings.... that still have effect sizes that aren't clinically meaningful.
We even know about things with large effect sizes that are controlled by a few variables, that are rare enough that it's hard to find the patients that they apply to.
And, of course, correlation ain't causation: if we used your system and found a bunch of things, we'd still need to figure out how to reduce them down enough to be something we could test before advising practitioners to do things a certain way.
On the other hand, RCTs gather relatively well-conditioned data. And we have enough study power across them that we end up with a lot of statistically significant, real findings.... that still have effect sizes that aren't clinically meaningful.
We even know about things with large effect sizes that are controlled by a few variables, that are rare enough that it's hard to find the patients that they apply to.
And, of course, correlation ain't causation: if we used your system and found a bunch of things, we'd still need to figure out how to reduce them down enough to be something we could test before advising practitioners to do things a certain way.