> the relative reduction is less impactful (e.g. 4% to 3.28%

That's also an 18% reduction

I think what he means is a reduction of 18% based on 4% is way less than 18% based on 80%.

Percents of percents always felt kludgey.

Log probabilities (like decibans) unify this to say there is a -0.86 dB risk reduction for everybody.

https://rationalnumbers.james-kay.com/?p=306

It makes the math of combining risks easier and works the same even if we're operating near 99.999% or 0.0001%

That’s exactly my point.

If someone is high risk, say 20%, then an 18% drop from that is 14.4%. That may justify picking up caffeine.

But if you’re otherwise healthy, picking up caffeine has diminishing returns, and the downsides may not be worth it.