1
$\begingroup$

I would like someone experienced to look at my KM curve. I have categorized a continuous clinical variable into 2 groups based on its median. The Log rank test is significant (p=0.0052).

I have then made models using Cox regression (PH assumptions checked) with known confounders. For the Cox regression, I treated this variable as continuous.

Can anyone tell me if the blue curve is somehow distorting the picture? The survival times distribution could influence this, correct? I plan to present adjusted survival curves based on the PROC PHREG (Cox regression) but I want to know how this raw survival curve stands.

KM curve using PROC LIFETEST

$\endgroup$
2
$\begingroup$

Your plan to base your analysis on a Cox model with the covariate treated as continuous makes a lot of sense, if both the proportional hazard (PH) assumption and linearity in the covariate hold true. This Cox multiple regression analysis is where you need the power that incorporating all the cases provides.

In that situation your Kaplan-Meier (KM) curves are mainly for display purposes, not for statistical analysis. All the log-rank test for the KM curve comparison tells you is the single-predictor relation of a binned continuous variable to outcome. That's of much less importance than the relation of that continuous variable to outcome in a Cox model that incorporates information about other covariates. It would be quite possible for this type of (single, binned-predictor) log-rank test to be insignificant statistically while the continuous predictor is still significantly related to outcome in a more complete Cox model.

So there would be nothing wrong with simply truncating the KM display at, say, 10 or 15 years. You could still base the log-rank test, if you need to show it, on all cases. Good practice is to present, under the KM curve, a table of cases still at risk in each subgroup at selected times. Among other things, that information will tell the reader how many cases with censoring or events at later times have been omitted from the plot. Those with experience looking at KM plots understand that the curves start to look ugly at late time points when few cases are still at risk, and they will not be misled by this type of presentation.

$\endgroup$
0
$\begingroup$

It seems that there are a few observations at the very end which took longer times for the blue curve to drop. If you subset your data to survival years greater than 15 and see the counts for the each drop that would explain why the blue curve is distorting

$\endgroup$
  • $\begingroup$ Thank you Ramu. I have thought about subsetting the data. But actually, doing so would lose me some power. Given that Proportionality is conserved I'm wondering if it is necessary to even consider truncating as you suggest $\endgroup$ – mindhabits Mar 15 at 22:06

Your Answer

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

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