10
$\begingroup$

In the context of a Cox regression, I recently heard someone say that "the hazard ratio of 0.70 between the two treatment groups indicates that, during the whole followup time, of 100 persons in the control group who have an event only 70 have an event in the intervention group." Is this a valid simplification of an HR or are there other non-technical interpretations?

$\endgroup$
5
  • 1
    $\begingroup$ That isn't sensible English. Looks like some kind of typo. $\endgroup$
    – Peter Flom
    Commented Jun 11 at 11:52
  • 1
    $\begingroup$ @PeterFlom The quoted sentence reads as cromulent English to me. (I would prefer the word "entire followup period" to "whole followup time" but the quoted sentence is understable with "whole followup time"). $\endgroup$
    – Alexis
    Commented Jun 12 at 16:28
  • 1
    $\begingroup$ @Alexis Thanks for teaching me a new word! $\endgroup$
    – Peter Flom
    Commented Jun 12 at 16:36
  • $\begingroup$ The interpreter may rather be interested in number-needed-to-treat (NNT). $\endgroup$
    – AdamO
    Commented Jun 12 at 17:04
  • $\begingroup$ Sorry, English is not my mother tongue. I indeed meant what is discussed above (entire followup period). $\endgroup$
    – Survival
    Commented Jun 14 at 10:43

2 Answers 2

16
$\begingroup$

A simple explanation of the hazard ratio is:

At any given instant, people in group A are XXX as likely to have the event as people in group B.

"Over the whole of the study period" is not right, because each person can have a different study period, due to either censoring (which could be left or right) or the event occurring.

$\endgroup$
4
  • 6
    $\begingroup$ The obvious example being a study (like cancer treatment) that ran for so long that everyone dies in both groups. Event ratio would obviously be 1:1. $\endgroup$ Commented Jun 11 at 12:28
  • $\begingroup$ +1 The quote in the question works sort of Ok as a simplified explanation, but as you point out obscures Cox PH's many assumptions. $\endgroup$
    – Alexis
    Commented Jun 12 at 16:50
  • $\begingroup$ The problem with this interpretation is the confusion between incidence and cumulative incidence. "Have the event" is ambiguous, whereas "the event occurs" is more to the meaning. $\endgroup$
    – AdamO
    Commented Jun 12 at 16:55
  • $\begingroup$ Thanks, that does make sense to me. If the hazard between the groups was proportional, then the HR applies at each timepoint t (See Frank‘s answer below). So why wouldnt this definition apply to the entire followup time? Does the noncollapsibility of HRs play a role in this as the risk set is reduced with ongoing time? $\endgroup$
    – Survival
    Commented Jun 14 at 10:40
14
$\begingroup$

We need to start with a premise. Suppose that either the HR is constant (so that what’s below applies to any t) or that we are speaking at the hazard ratio at a specific time t if the HR is not constant. First think of discrete time. The group B to group A HR is the B:A ratio of the probabilities of failing at discrete time t, given the subject had not failed before t. Now consider continuous time. The B:A HR is the ratio of two limits (hazard rates) as the time interval shrinks to zero. The hazard rate is the limit as $\delta$ approaches zero of the ratio of the probability of failing within $(t, t+\delta)$ given not failing before $t$, to $\delta$.

$\endgroup$
2
  • 1
    $\begingroup$ Not to be rude, but you seem to not given much notice to the "layman's terms" part of the question. $\endgroup$ Commented Jun 15 at 21:30
  • 1
    $\begingroup$ The only thing I can think of expanding is this: A hazard function expresses the level of instantaneous risk and this varies over time. These instantaneous risk are called ‘force of mortality’ in demography and represent the instantaneous chance of suffering the event at a given time given the event has not occurred before that time. $\endgroup$ Commented Jun 16 at 12:37

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.