This paper on #COVID19 #thrombosis has gained a lot of attention.

https://doi.org/10.1016/j.thromres.2020.04.013

The paper (n=184) reports a cumulative thrombosis incidence of 31% in hospitalized patients, but in Table 3 there are only 31 events (31/184) = 17%.

Why the large discrepancy???

1/

The reason is that a fairly large number of patients (n=45) were censored (i.e., dropped out of the sample before having an event).

If we are willing to assume that no extra thromboses occurred in censored patients, then 17% would be correct.
2/

The 31% estimate is from (I’m assuming) a Kaplan Meir cumulative incidence function. KM assumes that censored patients have the same risk as uncensored patients – in effect, assuming that some of the censored patients go on to have a thrombosis.

3/

The earlier that censorings occur, the larger the discrepancy between crude incidence and KM incidence.

So, which is the better estimate? 17% or 31%?

Well, that depends – why are patients censored? And, what is a reasonable assumption about thrombosis risk after censoring.

22 patients were censored because of discharge --> one could argue risk is lower in these patients.

5/

But wait, there’s more.

23 patients died. If deaths are treated as censored, does it make sense to assume that some of the dead patients go on to develop thrombosis?

Perhaps not.

So, how to account for the deaths?

6/

This is a scenario where a competing risk model should be used. In particular, the subdistribution hazard (i.e. Fine & Gray) model, which does not assume that patients remain at risk after they die.

See here for a review:
https://cjasn.asnjournals.org/content/12/7/1181.long

7/

How would the Fine & Gray estimate compare to the standard KM estimate?

It would be lower, because as mentioned, it doesn’t assume additional events occur after patients die.

8/

So, if you’ve made it this far, what do think is the best estimate for thrombosis incidence in this population?