I would like to get explained the following situations in coxph (library(survival)
)
- The Concordance, what is a good Concordance?
- What does it mean
robust= 2.61 p=0.1
in alogrank p=0.006
? - what is a good rsq?
(summary(coxph.model) $ rsq
- How to interpret
cox.zph (coxph.model)
– the output has, for every variable in the multivariable coxph.model achisq
,df
andp value
. Also, in addition to the variables in the multivariable model, there is a GLOBAL value.
This would be a potential output (I've tricked the name of the variables due to confidenciality) – I think it would be helpful to have the interpretation of this particular output as a guidance, if possible.
thank you!
Best Answer
Point 1. Quoting from Section 20.10 of Frank Harrell's Regression Modeling Strategies:
So a concordance of 0.5 is what you get if a model can't distinguish survival times at all. What's "good" above that depends on the nature of the study.
Point 2. Standard significance tests assume that observations are independent. Your use of an
id
variable indicates that some individuals (or groups of individuals with the sameid
value) contributed to multiple observations. Perhaps a single individual could experience more than 1 event. The "robust" standard error estimate takes that lack of independence into account, generally leading to wider confidence intervals and higher p-values. As the output from the summary says:Point 3. See the answer to Point 1. What's good depends on the nature of the study. I find concordance and measures of model validation and calibration to be more useful than $R^2$ values. I strongly recommend learning to use the tools in Frank Harrell's
rms
package to evaluate model quality.Point 4. This is covered at the end of Chapter 3 (Section 3.5.2, "Score tests") of the main survival vignette.
for individual predictors or for the model as a whole (GLOBAL). A low p-value indicates evidence against PH. In your case, it looks like
age
might violate PH. You might fix that by modeling age flexibly with a spline (e.g.,rcs()
in therms
package), as incorrect specification of the functional form of a continuous predictor can show up as an apparent violation of PH.