Change in C-statistic (with confidence intervals) on imputed survival data

[Brendon Neuen]

Hi everyone,

I am trying to calculate C-statistics (and change in C-statistics) with associated confidence intervals for my Cox regression model.

I have done so using the somersd package, developed by Roger Newson, for the complete case analysis. (Details here https://journals.sagepub.com/doi/pdf...867X1001000303)

After running the Cox model (stcox var1 var2 var3 etc), I used the "predict" command to calculate a hazard ratio,and then derived an inverse hazard ratio, which we expect to be a positive predictor (as suggested in the instructions in the link).

. stcox var1 var2 var3
. predict hr
. generate invhr=1/hr

I am trying to do the same for my full dataset with multiple imputation. After running Cox regession on the full imputed dataset (mi estimate, saving(mi_est) hr: stcox var1 var2 var3 etc), I used the "mi predict" command (mi predict hr using mi_est).

However this calculates linear prediction, rather than the relative hazards. Does anyone have any suggestions or solutions to this problem?

I also appreciate the importance of using a training and test set, as suggested in the above document by Dr Newson.

Many thanks,
Dr Brendon Neuen
The George Institute for Global Health

[Jeph Herrin]

You probably want to do this on each imputed dataset and pool the results. See https://www.stata.com/support/faqs/s...-imputed-data/, search for the section on -eroctab-

hth,
Jeph

[Brendon Neuen]

Thanks Jeph, I will check it out