Dear Joseph,

I guess You use R-package pPOC.

If You want perform ROC-analysis for multivariable logistic regression  model, You should create model like

GLM

Thanks Olga - much appreciated!

Hi Olga

Another couple of questions if I may.

Originally the models of interest were generated using coxph analysis, followed by MuMIn analysis to identify the best covariate combinations (evidenced by AICc). In the example above, you suggest using glm - would I need to use coxph here instead?

This also becomes more complicated since I'm using a combination of binomial and continuous variables, thus deciding which distribution to fit becomes a challenge!

Any advice you can provide would be gratefully received!

Thanks

Joe

Hi Joseph,

1 If You use Cox-regression - AUC depends on time.

Appropriate packages are  timeROC,survivalROC,risksetROC ,  survAUC.

2 "which distribution to fit becomes a challenge" -Cox-regression is nonparametric. We don't need to know the distribution.

3 If You want know how well is Your Cox-model - maybe concordance index would help ?(rcorr.cens (), library Hmisc )

My best wishes!

Olga

Hi Olga

thanks for the additional information. I'll check out the other packages you suggest. I gave the glm approach you suggested above a whirl - it seemed to work well (thanks!). Hopefully I can just tweak the code to incorporate the other tools you suggested.

RE checking model fit, I've been using an AICc approach via the MuMIn package.

Again, many thanks for your assistance with this.

Joe