Dear everyone,
I am performing a cox proportional hazard regression on survival, in a sample in which almost everyone dies in the follow up period. I am far from an expert, and want to be sure to thoroughly check the PH assumption. I can get some help from a statistician, but it takes 2 weeks til i can get an appointment (I will!)
I found 2 methods for checking the PH assumption that i can easily perform in SPSS: visually I can inspect stratified log minus log plots (and scatterplots of residuals for continuous variables). My question is about the statistical method: checking if the product of time and my variable becomes significant in the cox regression (if yes, not fulfilling PH assumption). I havenoticed that it is quite common to first make an univariable cox regression for each covariate. I have been reading on the subject but see that different methods are used when it comes to check individual Time dependent covariates.
Some people check the product of time*variable (T_COV) univariable in the cox-regression, others put both the T_COV and the original variable in the cox regression (example for age: T*Age and age would both be taken into the cox-regression). In the second method, one does noet acquire univariable cox regression for the T_COV.
Why is it important to me? There is one T_COV variable that is not significant in univariable cox- regression, but becomes significant in the regression with only the T_COV and the original variable that it is a product of.
I hope you want to give me your thoughts on the topic. I am extra happy with reading recommendations/reliable sources!
If you have other remarks, questions, or if my methods are all wrong: please comment!
Hi Josje,
Your question relates to time-dependent effects of covariates (such as age). Testing PH-assumption via T_COV tests only linear relationships over time, but the change over time may also follow other forms. A test is available that is based on scaled Schoenfeld residuals that test for violation of the PH-assumption. More importantly, these residuals can also be used to produce a plot that shows where in the time-axis the violation occurs and how problematic it is. As a practical example see Figure 1 in https://academic.oup.com/psychsocgerontology/article/67/6/765/612992. Violation of the PH assumption may also be an indication of non-linear effecy of the covariate, so you should check this also. There is useful guidance on these issues and several other assumptions of the Cox model in: Therneau & Grambsch (2000). Modeling Survival Data: Extending the Cox Model. New York: Springer.
Unfortunately SPSS is lagging in the diagnostics for the Cox model, so many of the useful features are not available in SPSS. Computing the residuals is possible via syntax, but it can be error-prone, and constructing the test can likewise be challenging. More advanced options are available in R (https://cran.r-project.org/). The book has good examples with code for R on how to perform diagnostics for your model.
As for your problem, you may need to discuss with your statistician about the approach to take for the analyses, because the useful solution is generally related to the study design, setting and data. Mostly in standard cox regression modeling scenarios I would use both the covariate and its time-interaction in the model.
Best regards,
Timo Törmäkangas
Gerontology Research Center
University of Jyväskylä
Finland
Hey Timo,
I have a similar question: what do I do if I used the tvc() option in order to test for the Cox-PH assumptions (interaction with time), and it returns a significant value for the covariate that first caused a violation of the Cox-PH assumption using the Schoenfeld residuals?
Kindest regards,
Okan
- Badges
- Science topic
- Similar topics
- Theoretical Physics
- Time