To continue . . . in statistics, survival analysis and mortality analysis are different. Survival measures time-to-event, and the relevant comparison is the hazard ratio. Survival curves are generated by log-rank or Kaplan-Meier functions. /1
Hazard ratios essentially measure the difference between the two curves, and express the difference in the probability of a subsequent event happening in each of the two groups. /2
Mortality measures the probability that a person in Group A or B is dead at a specified time point (on hospital discharge, 28 days after inclusion, 90d, 1 yr, etc.) The difference in mortality risk in two groups would be compared using chi-squared, Student's T, or similar. /3
You can have different survival curves but the same mortality. Think of looking at 100-yr mortality and survival between a group of heavy smokers and non-smokers. /4
The 100-yr mortality will be close to 100% in both groups. As Keynes said, "in the long run, we are all dead." /5
But survival will be greatly different. The time-to-event will be much longer in the non-smoker group. Stretched out over a long period of years or decades, looking at survival difference is meaningful, because patients care if the risk of death occurs within 15y or 50y. /6
But in the JACC anticoagulation for COVID article, the in-hospital mortality is basically identical. The survival curves are different. But the difference amounts to living for 7 days more with the same risk for death. /7
One way of interpreting this is that it takes 7 days longer for anticoagulated patients with COVID to die. Since most patients w/ COVID die in the hospital without family nearby, do we believe that this difference is likely to be important to patients? /8
Kaplan-Meier curves look very pretty, especially when you have different colors! Look at the difference between the colored curves! /9
Seriously, looking for survival difference is helpful, because it can permit getting a sense of whether mortality difference might be possible to observe, but survival analysis (esp K-M analysis) can be done with smaller N. /10
Thus, finding a survival difference in a RETROSPECTIVE, UNBLINDED cohort without estimates of residual confounding is hypothesis generating, but little more. /11
Anticoagulation, especially in the critically ill, has all kinds of risk. This paper reports a 50% increase (1.9% v 3.0%, P=0.2) in the risk of bleeding in the AC group. /12
As physicians, we have to use all the tools at our disposal, including good statistical interpretation, to weigh risk and benefit and make a recommendation about treatment. It is up to each physician to use professional judgment to make that calculation for each patient. /13
I see nothing in this report to incline a physician with good judgment to rush to anticoagulate in the absence of overt, clinically significant thromboembolism. /14
With luck, future RCTs with clear inclusion/exclusion criteria will help our judgment with better data. Until then, we must make treatment decisions with a high degree of uncertainty, like in response to so many questions. /15
First, do no harm. Don't just do something, stand there.
Hey, let's be careful out there.
Adding some stats people and some practitioners of #zentensivist care to get opinions.
@ADAlthousePhD @WalkeyAllan @stephensenn @EconTalker @msiuba @phlegmfighter @ogi_gajic
Hey, let's be careful out there.
Adding some stats people and some practitioners of #zentensivist care to get opinions.
@ADAlthousePhD @WalkeyAllan @stephensenn @EconTalker @msiuba @phlegmfighter @ogi_gajic
