I'm quite new to survival analysis, but would appreciate any advice on how to deal with the following finding in my data analysis/ results interpretation. I have been using R and have used a time-dependent Cox proportional hazards model to analyse risk factors for diabetic cats entering remission. I only have 46 cats in the study population and 13 remission events (patient recruitment is difficult in vet medicine.)

In univariate analysis, the following are significant: 1) BMI at enrollment (dichotomous, static covariate): HR 3.3 (95% CI 1.1-10) p= 0.04 2) Weight-loss (kg) (time-dependent covariate): HR 4.6; 95% CI 1.2 – 15.8; p=0.03

However, when entered into a multivariable model, both become non-significant (BMI has p-value 0.24 and weight change has p-value 0.17). I understand that this means neither remains significant in the presence of the other and wondered whether weight loss is the mediator for remission in cats with higher BMIs. I have looked into mediator analysis, but can find little on this when the mediator is a time-varying covariate. Does anyone think this is worth pursuing, or know whether it is possible using R?

Many thanks in advance.


1 Answer 1


(This is more of a comment than an answer, but I can't comment because I do not have enough reputation points).

Have you seen my question here? I have a similar situation in which I want to test for mediation using Cox proportional hazards models. My situation is slightly different because my mediator is not a time-varying covariate.

I do not believe weight loss could be a mediator, because weight loss would need to remain significant even in the presence of BMI. This is step 3 in Baron and Kenny's procedure for testing mediation. It seems more likely that BMI and weight loss are correlated and you are getting the observed effects because of multicollinearity issues.

  • $\begingroup$ Thanks for replying and directing me to Baron and Kenny's procedure! Having spent more time reading and looking at my data, I think it probably is the case that obesity and weight loss are multicollinear. I suspect you might have already seen it, but the following paper's additional files gives a detailed description of how the authors carried out mediation analysis on their Cox regression model: ncbi.nlm.nih.gov/pmc/articles/PMC3917547 $\endgroup$
    – cat.vet
    Commented May 8, 2016 at 14:53
  • $\begingroup$ I had not seen that paper. Thanks that is helpful as well! $\endgroup$
    – bigdataumd
    Commented May 9, 2016 at 19:34

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.