I'm on a "PCSK9" inhibitor (I remember the spelling by thinking of "Baron von Esskay", the cartoon sausage shill of my youth. Always confused with KPCMS, Kodak's Precision Color Management System, which I used to support).
Short version: bi-weekly injectable limits production of cholesterol in your bod, and in so doing keeps the bad cholesterol way down. Further than statins, which I'm unable to tolerate.
Cardiologists and GPs are SUPER-JAZZED about this prospect. But in the RCT info (FOURIER trial, with my PCSK9 agent), you find:
---
Composite of cardiovascular death, MI, stroke, hospitalization for unstable angina, or coronary revascularization.
9.8% vs 11.3% in treatment vs placebo arms
HR 0.85, 95% CI: 0.79–0.92, p <0.001
---
So, about 1.5% difference in heart/stroke trouble in a population of ~27,500 willing jamokes over about 2.2 years.
Good, but, I'm not sure I'd set off the fireworks display over this, given the cost. Sure, I'd like not to suffer a stroke or MI, but wow it's expensive to be a little less likely of it.
Is there a good reason for the SUPER-JAZZING that I'm missing here?
Well you'd have to dig into the trial in detail. But that's quite a short trial. I think doctors often expect that benefits compound over time, meaning that if you took it for years, the benefits might be even larger.
Beyond that, my general philosophy is "1.5% here, 1.5% there, pretty soon you're talking about real money".
Well, a micromort is a 1/million chance of death. If you think of micromorts for some kind of ongoing activity, then I guess you could convert them to a hazard *rate*, and thereby get some kind of hazard ratio out of them? For example, averaged across all ages, people in rich countries apparently experience around 20 micromorts of risk per day. And skydiving apparently amounts to 8 micromorts of risk per jump. So if you went skydiving every day that might mean a hazard ratio of ≈ 28/20 = 1.4 meaning a change in life expectancy of around ln(1/1.4) * 12.87 ≈ -4.33 years.
But... the assumption of a constant hazard ratio is extremely inappropriate for skydiving, since it's an additive factor, not a multiplicative factor. So I guess that's my answer? Hazard ratios are better when risks multiply, micromorts are better when risks add. The former probably better for "medical" stuff, while the latter is probably better for things like skydiving.
I have to admit I reached the end of the article with a "I kinda got some of it" mental model of what was going on, but this viz + re-reading the parts I got lost on more critically got me most of the way there. Interesting stuff, particularly as I'm considering whether or not I should start a statin young.
I'd also really like to know if Claude looked at it! The simulation is remarkably similar in some ways. Is there any way for you to tell?
Incidentally, I've found myself skipping certain technical details these days, on the theory that once you get too far from the main path, people will be better served by asking an AI to fill in the blanks
wait, does this push you toward plans with explicit expiration dates? like, if a plan survives six months only because nobody re-ran the hazard math, that feels like fake confidence.
I'm on a "PCSK9" inhibitor (I remember the spelling by thinking of "Baron von Esskay", the cartoon sausage shill of my youth. Always confused with KPCMS, Kodak's Precision Color Management System, which I used to support).
Short version: bi-weekly injectable limits production of cholesterol in your bod, and in so doing keeps the bad cholesterol way down. Further than statins, which I'm unable to tolerate.
Cardiologists and GPs are SUPER-JAZZED about this prospect. But in the RCT info (FOURIER trial, with my PCSK9 agent), you find:
---
Composite of cardiovascular death, MI, stroke, hospitalization for unstable angina, or coronary revascularization.
9.8% vs 11.3% in treatment vs placebo arms
HR 0.85, 95% CI: 0.79–0.92, p <0.001
---
So, about 1.5% difference in heart/stroke trouble in a population of ~27,500 willing jamokes over about 2.2 years.
Good, but, I'm not sure I'd set off the fireworks display over this, given the cost. Sure, I'd like not to suffer a stroke or MI, but wow it's expensive to be a little less likely of it.
Is there a good reason for the SUPER-JAZZING that I'm missing here?
Well you'd have to dig into the trial in detail. But that's quite a short trial. I think doctors often expect that benefits compound over time, meaning that if you took it for years, the benefits might be even larger.
Beyond that, my general philosophy is "1.5% here, 1.5% there, pretty soon you're talking about real money".
I'll make the argument that everyone's "life expectancy at death" is 0 more years.
This was so cool, although I'm surprised micro-morts/micro-lives weren't mentioned. Isn't there a similar idea going on there?
Well, a micromort is a 1/million chance of death. If you think of micromorts for some kind of ongoing activity, then I guess you could convert them to a hazard *rate*, and thereby get some kind of hazard ratio out of them? For example, averaged across all ages, people in rich countries apparently experience around 20 micromorts of risk per day. And skydiving apparently amounts to 8 micromorts of risk per jump. So if you went skydiving every day that might mean a hazard ratio of ≈ 28/20 = 1.4 meaning a change in life expectancy of around ln(1/1.4) * 12.87 ≈ -4.33 years.
But... the assumption of a constant hazard ratio is extremely inappropriate for skydiving, since it's an additive factor, not a multiplicative factor. So I guess that's my answer? Hazard ratios are better when risks multiply, micromorts are better when risks add. The former probably better for "medical" stuff, while the latter is probably better for things like skydiving.
Neat. Now I'm curious if we can get variance estimates for these hazard ratios and adjust for the opimizers curse:
https://titotal.substack.com/p/the-best-cause-will-disappoint-you
I thought Claude (Opus 4.8 - got hit by the Fable classifier) did a good job making nice visualizations for this and summarizing them well: https://claude.ai/public/artifacts/dffb1a6c-a558-46f3-81c4-39a4edcf05ae
I have to admit I reached the end of the article with a "I kinda got some of it" mental model of what was going on, but this viz + re-reading the parts I got lost on more critically got me most of the way there. Interesting stuff, particularly as I'm considering whether or not I should start a statin young.
Did you try the much uglier but more comprehensive simulation DeepSeek+dynomight made? https://dynomight.net/img/hazard-ratios/simulator.html
I'd also really like to know if Claude looked at it! The simulation is remarkably similar in some ways. Is there any way for you to tell?
Incidentally, I've found myself skipping certain technical details these days, on the theory that once you get too far from the main path, people will be better served by asking an AI to fill in the blanks
I don't see it accessing that URL in its reasoning trace: https://claude.ai/share/8eeb8a2b-7873-4cfe-8678-917ea2fc839d
Nice. I'm going to want to chew on this a bit.
Pre-registered prediction: This will be one of the least-popular things I've ever written. ♡
wait, does this push you toward plans with explicit expiration dates? like, if a plan survives six months only because nobody re-ran the hazard math, that feels like fake confidence.