Thread by @LucyStats, A paper recently came out in NEJM on COVID-19 patients using remdesivir: [...]

A paper recently came out in NEJM on COVID-19 patients using remdesivir:

https://www.nejm.org/doi/full/10.1056/NEJMoa2007016

Bonovas & @piov1984 wrote a great critique pointing out that the analysis had a flaw!

https://www.nejm.org/doi/full/10.1056/NEJMc2015312?source=nejmtwitter&medium=organic-social

This thread is on *competing risks* & why it is so important! 1/13

Compassionate Use of Remdesivir in Covid-19 | NEJM

Correspondence from The New England Journal of Medicine — Compassionate Use of Remdesivir in Covid-19

https://www.nejm.org/doi/full/10.1056/NEJMoa2007016

First let's talk about the *question*

I believe the authors are interested in telling us about

clinical improvement in this cohort of patients taking remdesivir, in particular they want to estimate the cumulative incidence of clinical improvement by 28 days.

2/13

They define "clinical improvement" as:

being discharged alive or

having a decrease of 2 points or more in a 6-level ordinal scale of oxygen support:

ECMO

Mechanical ventilation

NIPPV

High-flow oxygen

Low-flow oxygen

Ambient air

3/13

SO we want to quantify the time to this event. To do this, we need to classify each patient as

having the event (improving) and record

WHEN it happened, or

we will censor them, meaning we didn't observe the event and record the last time

we saw the patient
4/13

I spent some time trying to recreate the analysis by pulling from their Figure 2, and I wasn’t quite about to do it. So I’ve painstakingly pulled every number from Figure 3A

. Here's what I got in case you'd like to follow along in #rstats

https://livefreeordichotomize.com/2020/05/22/survival-model-detective-2/

5/13

I spent some time trying to recreate the analysis by pulling from their Figure 2, and I wasn’t quite about to do it. So I’ve painstakingly pulled every number from Figure 3A . Here's what I got in case you'd like to follow along in #rstats https://livefreeordichotomize.com/2020/05/22/survival-model-detective-2/5/13

And here is the plot, yielding an estimated cumulative incidence of ~84% at 28 days. (replicating Figure 3A from the paper). Ok, cool what's the problem? Remember how we classified each patient as

having the event (improving) or

censoring them? The

is crucial here!
6/13

And here is the plot, yielding an estimated cumulative incidence of ~84% at 28 days. (replicating Figure 3A from the paper). Ok, cool what's the problem? Remember how we classified each patient as having the event (improving) or censoring them? The is crucial here!6/13

A key assumption is that the censoring is *non-informative* the patient is followed for a certain time, never has the event, & then is no longer followed. Our best guess is we know they didn’t have the event up until the last day we saw them, so we censor them on that day.
7/13

In order for the assumptions to be appropriately met, it must be the case that patients who have been censored are just as likely to have the event as those who are still being followed in the study. Let’s pull up a Figure to see the trajectory of all patients.
8/13

The black squares for 7 patients indicate they died! If someone dies we know that they are not going to improve later. This is not non-informative censoring!! Luckily there is a very straightforward way to deal with this in statistics: competing risks!
9/13

(Side note, I think this plot is very cool, if you'd like to learn how to make it using #rstats, check out this post)

https://livefreeordichotomize.com/2020/05/21/survival-model-detective-1/
10/13

Ok, competing risks! In a competing risk analysis, we can separate out the the death outcome from the remaining censored outcomes. We can then appropriately estimate the cumulative incidence of improving. I've fixed the data to account for the 7 deaths as a *competing risk*
11/13

Now we can plot the outcomes separately. Under this corrected framework, the estimated cumulative incidence of clinical improvement by day 28 is 74%
12/13

Competing risk analyses are so important and come up all the time in medical statistics.

Here are some great resources:

@zabormetrics has a great tutorial on doing these in #rstats: https://www.emilyzabor.com/tutorials/survival_analysis_in_r_tutorial.html#cumulative_incidence_for_competing_risks

I like quick explainer: https://pubmed.ncbi.nlm.nih.gov/15305188/
13/13

A Note on Competing Risks in Survival Data Analysis - PubMed

Survival analysis encompasses investigation of time to event data. In most clinical studies, estimating the cumulative incidence function (or the probability of experiencing an event by a given time)...

https://www.emilyzabor.com/tutorials/survival_analysis_in_r_tutorial.html#cumulative_incidence_for_competing_risks

You can find a summary of this explainer on competing risks here:

https://livefreeordichotomize.com/2020/05/22/survival-model-detective-2/

Take aways: Learn about proper survival analyses or befriend a statistician to help you!

14/13

You can follow @LucyStats.

Tip: mention @twtextapp on a Twitter thread with the keyword “unroll” to get a link to it.

Latest Threads Unrolled: