Not gonna lie, I’m still wrapping my head around RAR (Response-Adaptive Randomization). Anyone have a simple breakdown? #idsky - ThreadSky

absteward.bsky.social • 16 days ago

Not gonna lie, I’m still wrapping my head around RAR (Response-Adaptive Randomization). Anyone have a simple breakdown? #idsky

Reposted from Antibiotic Steward 🅱️C🆔🅿️🌟

🔥REMAP-CAP RCT🔥
Among patients admitted to the ICU with severe non-pandemic CAP & either cardiovascular or respiratory failure there was a low probability that a 7d fixed-duration course of hydrocortisone improved 90d mortality @pulmcrit.bsky.social #EMIMCC #IDSky
link.springer.com/article/10.1...

Comments

jonathanrydermd.bsky.social•16 days ago

@steventong.bsky.social might need your help here lol

tai-md.com•16 days ago

3-part post please. My Attention span drops to nothing after that.

kertviele.bsky.social•16 days ago

I posted an explainer, but it's over 3 posts. If you want the quick version...during a trial allocate more patients to better performing arms. Compared to fixed randomization, this treats patients in the trial better, and has a higher chance of getting the correct answer (identify best arm, etc)

tai-md.com•16 days ago

Interpretation: It’s legit.

jonathanrydermd.bsky.social•16 days ago

It sounds cool, but in this trial steroids looked worse, so how did so many get assigned to that arm?

kertviele.bsky.social•16 days ago

Basically three things happened simultaneously.

1) There was a data error in an interim (arms were labelled incorrectly, this is in the paper). This skewed the allocation probabilities the wrong direction. Ordinarily this could be corrected at the next interim, but.....

drandrekalil.bsky.social•16 days ago

https://pubmed.ncbi.nlm.nih.gov/32222766/

absteward.bsky.social•16 days ago

Thanks for sharing #idsky

kertviele.bsky.social•16 days ago

So this is one of the frustrating papers. I agree with the conclusions that you shouldn't do RAR in two arm (control/treatment) settings, and that you shouldn't do 1980 style play the winner. But modern multiple arm RAR doesn't do these things, and none of the modern methods are referenced.

drmichaelmarks.bsky.social•16 days ago

I found reading your simulation papers particularly useful here. In general I find a well done simulation paper makes some of these fairly non-intuitive concepts more concrete and allows a trialist but not statistician (like myself) to have a much better grasp of the issues. More please!

drmichaelmarks.bsky.social•16 days ago

One issue am interested in (& has come up in discussions about some our trials - SNAP, etc) is handling centres only randomising to some arms (As happened in REMAP-CAP). Can see why clinicians (like me) like this but worry about what it says about ethical equipoise & statistical efficiency/accuracy.

drmichaelmarks.bsky.social•16 days ago

Would intuitively think generates interaction between site effects & estimating outcomes which might not be fully adjusted for just by having a 'site' term in the model as that effect is more than just arm preferences? Presumably are simulations showing how performs under different scenarios...

kertviele.bsky.social•16 days ago

Why is multiple arm RAR different than two arm? If you want to estimate a treatment effect, you need good estimates on both treatment and control. In two arm RAR, any extra allocation to treatment comes at a cost of decreased control allocation. You lose power....but....

kertviele.bsky.social•16 days ago

In multiple arm RAR, you may be most interested in finding the best arm and comparing to control. If you keep control allocation constant (nothing lost) and do RAR among the active arms, you will tend to get more patients on the best arm, and get more power in comparing best arm to control

kertviele.bsky.social•16 days ago

If you are interested in comparing ALL active arms to control, then you might not want RAR (the best active arm gets more patients at the cost of lesser arms getting fewer). Different problem, different solution.

Comments

Posting Rules

Reply