Hacker News new | ask | show | jobs
by abecedarius 1219 days ago
Good question: it turns out I didn't know there's a technical difference between "hazard ratio" and "relative risk". https://en.wikipedia.org/wiki/Hazard_ratio

But the basic idea is that in the samples under the two different conditions, the ratio of the observed frequencies is 1.3 or 1.8. E.g. it'd be 1.8 if in the fasting sample they saw 3.6% bad outcome, while in the nonfasting sample they saw 2.0% bad outcome.
1 comments
VernorVintage 1218 days ago
Okay, wow so HR 1.8 literally does mean that fasters are nearly twice as likely to die over a given period as non-faster. But the death rate for fasters and nonfasters can still be pretty low, for example 2% a year vs 1% a year.

Since the risk is low already, an individual might be indifferent to a 1% or 2% death rate - from an individual perspective the two are even statistically indistinguishable (they can't exactly die more than once!)

Okay, I'm a little less freaked out by the seeming /doubling/ of death rate, however, it's still 2x increase in death rate. Still chilling!

Now I suppose that if we assume the death rate is constant for everyone at all times - or use a more realistic actuarial curve - then we can even transform the HR factor into an average effect on lifespan for fasters vs nonfaster.

Then we could say "ah, nonfasters will die on average 4 years sooner than nonfasters". I think that would help contextualize the numbers even better than relative death rates (which are hard for individuals to reason about).

Are you familiar with how we might convert death HR into an expected lifespan delta for the average person? Would that analysis actually hold any meaningful truth?

link
abecedarius 1218 days ago
> Are you familiar with how we might convert death HR into an expected lifespan delta for the average person?

You'd have to integrate over lifespan using an actuarial table, I assume.

This figure of 1.8 wasn't for all-cause mortality, but rather for heart-disease deaths among breakfast-skippers. I wouldn't worry (much) for a bunch of related reasons:

- The more complexity you need to specify a statistic, the more evidence you need to overcome a presumption it's an artifact. (Some of this caveat shows up in how they report it ("multivariable-adjusted", "95% confidence interval"), but these frequentist-statistics methods don't really account for all the real-world reasons for doubt.)

- This study was observational, not interventional. It's one study. Observational studies, particularly in epidemiology, particularly for diet, are often misleading.

- What'd be the mechanism making skipping breakfast have this big an effect? There could well be a good answer to that, but it doesn't spring to mind for me. In other words, what was your prior on this hypothesis, before this paper?

- I haven't read the paper beyond the abstract linked to, but https://news.ycombinator.com/item?id=34883541 lists some potential problems.

- The process that brings a particular medical paper to your attention, unfortunately, is not much like one that optimizes for the most significant updates to the probability distribution you should have over hypotheses. Unless you have a much better news feed than most of us, anyway.

I think of a result like this as a little bit of evidence to keep in mind as you learn more. It's not making me do anything different right now. That might seem harsh about a study involving "185,398 person-years of follow-up period" -- that must've taken a fuck-ton of work -- but it seems to be the world we're in.

link