Thread by @tmorris_mrc

Hot off the press: I’m excited to share the results of The KMunicate study with you!

This one is on your very favourite topic: Kaplan–Meier* plots.
*Yes, with an en-dash
1/
bmjopen.bmj.com/content/9/9/e0…

You all know what a Kaplan–Meier plot is? Yeah you do. You use them with time-to event data when you want an estimate of the cumulative proportion ‘surviving’ over time, and you plot it.
Here’s one from the RT01 trial.
2/

This project started off with @mattsydes, @TheJarvisC1 and I ranting about the poor quality of Kaplan–Meier plots we see in RCT publications. We decided to discuss what we thought people *should* be doing.
3/

The two main things we griped about?
1. One important way that people use K–M plots in trials is for description. So it would be good to know what is happening to patients over time, right?
2. People seem to easily forget about uncertainty when reading K–M plots
4/

A few more people joined us (all co-authors) and we came up with a few proposals to try and address these issues.
The group had different opinions about which were most useful. Someone – I forget who, but it wasn’t me – suggested we run a survey.
5/

So we worked it up and worked really hard to publicise it all over the place (a-whole-nother thread) and got 1,174 responses in 6 weeks 🥳

Probably thanks to using this desaturated photo of a printed K–M plot in promoting it

I’d like to say now how grateful we are to every one of you who took the time to complete the survey (even the person who used the free-text fields for weird insults).
7/

If you want to see each of our proposals with a description, take a look at the article’s supplementary file:
bmjopen.bmj.com/content/9/9/e0…

So who got involved and what did they think of our proposals?
Here is ‘table 1’* describing participants
*actually not a table (thanks Cleveland?)
9/

Well, I haven't shown you the proposals themselves (you'll have to read the paper) in the thread but here is the main results figure summarising how people ranked proposals

People liked:
1. A clear table showing the status of participants (at-risk, censored or evented) over time
2. confidence intervals to describe uncertainty about the estimates
Like this
11/

Here is the link again: bmjopen.bmj.com/content/9/9/e0…
That’s it!
Oh, except for a supplementary thread
S1/

Here’s an interesting note on the ‘extended risktable’: the usual* table of the numbers at risk beneath a K–M plot displays the number at risk the instant before time t, so that anyone who fails at exactly time t is included in this number.
*@Stata implementation, at least
S2/

Our extended table required presenting the number at risk *at* time t, not the moment before, so that, at any time point, the numbers at-risk, censored and evented sum to n.
S3/

Clearly the extended table doesn’t need all three rows, since if you know two you can work out the third, but seeing that each column sums to n is intended to reassure the reader.
S4/

Also, this extended risktable isn’t currently implemented in any software.
Fancy fame, fortune and frustrations? Then do go ahead and program it!
S5/

I’ll say something that bugged me about the survey and that seems to be common to all such surveys: there is no well-defined target population. The target is ‘anyone with enough of an interest to choose to participate’.
S6/

We will never know who didn’t see the survey but would have participated if only they had. This means we don’t have any concept of what sort of sample we got.
Any thoughts on this?
S7/