Cox model: c-index

[EQUIWDH]

Hello, I 'm using the package to calculate some real survival data, I find the cross-validation can only choose deviance on the test cohort to determine the support size. could you guys add the c_index principle for cox model's cross validation?

[Mamba413]

@oooo26 ，I think we can first work on classification with auc, and then, turn to this issue.

[oooo26]

No prob, I will take a look.

[Mamba413]

definition of C-index: https://statisticaloddsandends.wordpress.com/2019/10/26/what-is-harrells-c-index/.
I also recommend implementing C-index following some mature libraries like ranger: https://github.com/imbs-hl/ranger/blob/master/src/TreeSurvival.cpp.

[EQUIWDH]

It seems the whole constructure do not provide a parameter space for time vector when the situation is censored ,because the T1 *y for cox model only pass the censoring indicator sorted by time column to calculate loss enough. However, If hope to calculate Harrel's c-index for censored data especially when $T_i = T_j$ shows up. The time shown in bess_base.py must be passed for the initialization of abessCox in AlgorithmGLM.h. I find other file defaults y is a vector so I guess I can not directly regard it as a n*2 matrix to pass the parameter. Hence, the other way is adding an initialization parameter for time. could u take a look how to do it because I think I must update many other files and it's a little bit tough.

[Mamba413]

@oooo26 ，what about add a parameter for API?

[oooo26]

Actually modifying the API is easy, but as you discussed, I think the trouble is how to use time in CV, e.g.:

• If stacking time with y, you need to change the Cox model into multiple one;
• If storing time into Data, you need to modify CV splitting process and pass it to score computation;
• If adding time as a seperated/subclass variable, more functions need to be changed and it would be annoying;

I would recommend the second way. But indeed, each of them requires updates on many files, so if there is a better idea, please do it.