6

In a dataset with survival event, I calculated Harrell C-index from three different models. Furthermore, I calculated the 95% C.I. for the three different models.

So the next question is to compare the discrimination ability of the three models.

Could anyone kindly suggest the method to compare the Harrell C-index?

mpiktas
  • 33,140
  • 5
  • 82
  • 138
Yao Zhu
  • 111
  • 1
  • 1
  • 7

2 Answers2

4

Harrell would advise that you NOT do so:

How to do ROC-analysis in R with a Cox model

Doing model comparison with LR statistics is more powerful than using methods that depend on an asymptotic distribution of the C-index.

Community
  • 1
DWin
  • 7,005
  • 17
  • 32
  • Thanks DWin. Actually I have performed LR test to test the predictive adequacy index of the 3 model (vs. combined). The confusing point is I found the model with the highest C-index but have moderate adequacy index. – Yao Zhu Oct 28 '11 at 07:45
  • Or another alternative is: using bootstrap, I calculate the 3 C-index in every resampled population. Thus, I get the 95% CI of the C-index of the differences between model A/B/C. The one with 95% CI >0 is the "significant" better? – Yao Zhu Oct 28 '11 at 07:50
0

LR statistics are well suited for hierarchical models. In case of various variables in models I guess information criteria (Akaike - AIC or Bayes BIC) will answer which model is the best.

Piotr Bryniarski
  • 11
  • 2