1
$\begingroup$

A question that should be straightforward for statisticians but is puzzling me to the point I am nervous. I am doing a PhD in Epidemiology and am going to use Cox regression. The predictor is a histological score for how bad a cancer was at the moment it was surgically removed, and the outcome is patient's survival after surgery in months.

However, almost all previous biomedical papers on the same predictor and outcome include a strange sentence. They state something along the lines: "the endpoint evaluated was 5-year survival, and the methods were log-rank test and Cox proportional hazards regression". But: neither log-rank nor Cox regression care what the survival at a specific number of years was. Five-year survival is calculated in descriptive statistics because it is interesting for doctors, but that is not the real inferential statistical endpoint. The endpoint is continuous, a time-to-event variable in months (or any other time unit). So why the heck do ALL PAPERS state that the ENDPOINT is "5 year disease-free survival" and not "disease-free survival in months"?

At first I thought that they did a logistic regression (then you could use "5-year survival" as a 0-1 score, though it would be statistically sketchy). But they all use Cox in the end, yet they all state the endpoint like that in Introduction/Methods. Am I wrong? Am I missing something? Why would no reviewer correct that? I am getting crazy about it.

$\endgroup$
  • 1
    $\begingroup$ At the risk of offending some, a key reason is because doctors want the information they're used to seeing in the way they are used to seeing it, which is often not what is most to the point/informative. There are a lot of dinosaur methods in use in medical research because of this kind of conservatism, even though they've been replaced by better alternatives in the statistics literature. The obsession with $p$-values is also a symptom of this problem. $\endgroup$ – Marc Claesen Jul 14 '15 at 9:48
  • $\begingroup$ @MarcClaesen, you are right. I am a doctor. When starting my PhD I dreamed of improving biomedical epidemiology by "better statistics", only to realize that my colleagues/supervisors only wanted the "old" methods from the 70s (e.g. I have trouble getting imputation methods accepted). Same goes for ignoring effect size in favour of p-values. In this case, however, it is just plain wrong to say that the "endpoint" is 5-year survival. They should say "we estimated 5ys [in descriptive statistics] for categories of [var] and tested the predictive value of [var] for survival in months". $\endgroup$ – torwart Jul 23 '15 at 9:33

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.