2023 The Final Countdown, page-1642

Arent you the "one in twenth subgroups will show a significant p value" guy? and you espouse nuances of the scientific method as one of your highest values. Lol.

let me put a controversial idea to you for some "nuance": RCT is not necessarily the gold standard or always appropriate. Crazy I know but hear me out.

let's look at an extreme example to illustrate this issue: let's conduct an RCT of two children: one gets the treatment, the other placebo. It's slim pickings for this rare disease hence n is low. The two patients are a little different though: one is a 12 years old boy, has D grade severity of the disease, the other is 8 years old, female, and has C grade severity.

now we could compare these two for our trial, but there are obvious problems with this method. We need to understand how each of them perform with all of their personal characteristics taken into consideration. Unfortunately, although they have the same disease, the two kids are hardly comparable...

very fortunately, we have access to a large database of these kinds of patients. Let's call it, I dunno, the MAGIC cohort database, for instance. we're very lucky because we can match them by their characteristics:

100 day survival = b0 + b1*gradeSeverity + b2Age + b3Gender... etc

The model is statistically significant, and we can match each patient accordingly and appropriately. So what would we consider the gold standard in this situation: comparing the two RCT patients against each other, or against their respective *matched controls* in the database?

hopefully this illustrates why an RCT is not always Tennable for rare diseases. Sometimes we can't just rely on those idealism we picked up from that undergrad stats class that we barely scraped a pass in (we lost a lot of marks for our misunderstanding of type one and type two errors e.g. "one in twenty....").