0
$\begingroup$

I want to investigate significant predictors for disease free survival. For this, I did created Kaplan Meier curves for each possible risk factor. The significant predictors after Kaplan Meier analysis were included in a Cox regression analysis.

Tumor diameter (divided in 3 categories) was a significant predictor after Kaplan Meier analysis. Category 1 had a mean of 95.435, category 2 had a mean of 85.078, and category 3 had a mean of 64.382. The curves were plotted as expected: category 1 had the highest disease free survival, followed by category 2. Category 3 had the lowest disease free survival. After cox regression analysis however, tumor diameter remained a significant predictor, but the hazard ratio was 0.010 for group 2 compared to group 1.

Can anyone give an explanation on why the cox regression analysis predicts a significant higher disease free survival with increasing tumor diameter, while the kaplan meier analysis shows a decreased disease free survival with increasing tumor diameter?

$\endgroup$
1
  • $\begingroup$ Please provide more details of the data and the Cox model. Specify how many cases in each category, how many events, the summary of the Cox model output (coefficients and standard errors). A hazard ratio of 0.01 is awfully small, but without more details it will be impossible to see what's going on. Please provide this information by editing the question directly, as comments are easily overlooked and can get deleted. $\endgroup$
    – EdM
    Jul 8 '21 at 13:37
1
$\begingroup$

A simple explanation of the difference in your results is that the Kaplan-Meier estimator and the Cox model are different in nature. A Kaplan-Meier returns a non-parametric estimate of the survival function. To calculate it you simply need the number of individuals at risk and the number of those that exit at different points in time (eventually dealing with censored data). In your case, this is done for different sub-groups. A Cox model accounts for an array of covariates, simultaneously allowing for multiple predictors. Intuitively, they differ in the same way univariate and multivariate analyses do.

The answer you are looking for might lie among the independent variables you included (or did not include) in the Cox model. You could try to test and develop your model further to spell out any concerns (if deemed feasible/appropriate).

You could also try adding confidence intervals to your Kaplan-Meier plot (if you haven't done this yet), to check whether the mean differences you observe do actually hold over time.

$\endgroup$

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.