Not sure of the scientific background here but the most promising methods are platform and basket trials. Platforms allow you to test multiple therapies at once and focus on the best, basket trials allow you do to that in patient subgroups (ISPY 2 did both)

“A few subgroup analyses”

It’s a really bizarre business model to me. It seems aimed at SEEMING like our standard data analysis paradigms, but it’s much like me deciding to fit a linear regression and instead of telling you the coefficients, I instead give you a bunch of (X,Y) pairs generated from my fitted model.

None of this is properly “data”. At best it is a snapshot of a predictive distribution from some black box AI model fit. What you are estimating from such things is, at best, the AI model fit from the actual data (which the AI could just provide directly instead of “data” but often doesn’t).

Feels a bit like dominantly inherited Alzheimer’s. Genetic basis, somewhat predictable age of onset, etc.

From a modeling perspective we can have a debate. But in reality sites are far from random. Sponsors have preexisting arrangements with productive sites etc.

Other applications of hierarchical models here for borrowing across different regions of the world.

cdn.who.int/media/docs/d...

After a month I can’t remember what my own code does by reading the source (well at least it takes a while, and I don’t think I’m the worst coder). My help files are for me. If they help someone else too, great.

It's unclear the conclusions in that paper...that essentially the extra complexity isn't worthwhile...apply outside of the context of that paper, particularly the small sample sizes.

I don't know that I have seen a writeup specific to historical borrowing (too many things like that...). However the models are essentially identical to those in the basket trial literature like this paper.

www.tandfonline.com/doi/abs/10.1...

There are multiple versions. Some are simply a hierarchical model over studies. Others fit a two component mixture with one component having the current study independent of others. Others fit more complex models like dirichlet processes.

Of course, seeing people submit false AI references can be very informative about their quality control….we have already caught false references in potential FDA submissions.

But...but....I really like my
Bayesian Adaptive Design And Subject Simulator
software....

Agree with proper terms in the model the hybrid isn’t itself bad, although if you can individually randomize I would strongly recommend doing that from the start. Separate issue if you think one part of the hybrid is biased.

I can’t comment on the percentage of the industry that does such due diligence as standard practice, but we have done a lot of consulting for venture capitalists to vet what they are potentially investing in.

Ah you got me. It’s not that old (maybe 5 years) but my beard has turned grey in the meantime! It’s a good picture because my son was joking around in the background.

The right frequency is based on how often you take a good picture. Those simply could be 20 years apart for some of us.

Agree lots of problems implementing this. Amusingly this may mark the return of the travel agent, suddenly back to manipulate the AI that is trying to manipulate you.

Said like a true Bayesian