Thread by @tmorris_mrc

I've seen so many people reference/tweet @_MiguelHernan's seminal paper The Hazard of Hazard Ratios and, to my shame, only just got to reading it.
1/
ncbi.nlm.nih.gov/pmc/articles/P…

Aside: twitter has felt a bit less informative and more negative recently, and I miss people's fun threads/tweetorials so here's a thread.
2/

When people have mentioned this article, I feel like the message has been ‘the average HR has an inbuilt selection bias and cannot be interpreted causally’ – which it doesn’t claim at all!
3/

One actual argument is that the average HR may mask non-proportional hazards: we might feasibly see an HR>1 for a period then >1 for another period, which could average out as 1.
Worse, the average will depend on the distribution of follow-up time.
4/

So we use period-specific HRs instead, right?
This is where an argument about selection appears. Suppose we estimate period-specific hazard ratios for a binary randomised covariate…
5/

Hernan’s argument is that doing this for the first period would be fine, but at the second period, more people may have had an event in one group, leaving a relatively more frail group in the other (which bizarrely makes me think of the totally-unfair catch-up logic in 1080).
6/

The argument relies on the existence of some frailty – something other than the binary covariate that gives patients different hazards – absent from the models (this is absolutely the right thing to assume).
7/

He suggest some alternatives to a single average HR and period-specific HRs, including Kaplan–Meier estimates of survival at various follow-up times (which were at the time routinely omitted from observational analyses due to confounding)
8/

*ahem* plug plug ‘Have you seen the KMunicate study?!’ 😃
9/
bmjopen.bmj.com/content/9/9/e0…

Then, a nice strategy for estimating a method of estimating adjusted survival curves in observational studies (which I won't explain in detail, but based on producing standardised survival curves, marginalising over the observed covariate distribution)
10/

Hey #EpiTwitter, have you ever used this 👆 idea? How has it worked for you?
I’ve not really thought about it but it seems a nice idea.
11/

The end suggests that average HRs are still useful but we might estimate them over different follow-up times. I didn't understand this. Why not just estimate an HR that changes (smoothly) over time?
12/

So yeah, cool. Fizzle out to black.
(BTW I concentrated better than usual while reading this paper because I was trying to write a coherent thread!)
13/EoT