Thread by @hilaryarobbins

The all-cause mortality result in the NELSON #lungcancer screening trial has generated major debate.

How do we interpret this? Let's start with what we expected to observe.

@NEJM #lcsm #lungcancerscreening #EpiTwitter

https://twitter.com/VPrasadMDMPH/status/1222673150485262336

NELSON showed a 24% statistically significant reduction in lung cancer mortality with low-dose CT screening in men.

If the intervention (screening) does not lead to fatal harms, then we expect the rate of death from other (non-lung-cancer) causes to be the same in both arms.

Now, a large reduction in lung cancer mortality would translate to a small reduction in all-cause mortality, since lung cancer only causes a portion of deaths (see NLST).

NELSON did not have enough power to test whether the hazard ratio for all-cause mortality differed from 1.

Thus, the a priori expectation (given a reduction in lung cancer mortality) would be that the point estimate for the all-cause mortality hazard ratio would be less than 1, but with the confidence interval overlapping 1.

HR=1.01 was thus unexpected.

So, let's simulate 500,000 NELSON trials in which the observed LC mortality reduction occurred, and other-cause mortality occurred at same (mean) rate in both arms.

✳️ In 9% of such trials, we would expect to see an all-cause mortality hazard ratio of 1.0 or above. ✳️

This can't tell us why we observed the all-cause HR=1.01 in NELSON. It shows that it could conceivably be random chance.

My hope is that we will see more details published about causes of death in each arm of NELSON, along with details about harms from screening follow-up.