I've carried out a survival analysis. I have a dataset of live singleton deliveries over a few year period (~203, 000 deliveries, 1, 512 events). I'd like to be able to calculate the C-statistic/C-index for my proc phreg model. I have found various macros online to do this, and have them up and running. However, I am concerned that even though I take care of the clustering of children within mothers (mothers could have more than 1 live singleton birth in this three year period) using the covsandwich (aggregate) option, I'm not sure that the macros calculating C-index take clustering into account. I can see that they calculate the usable pairs, concordant and discordant pairs that go into the C-index calculation, but it's not clear they account for clustering. What would you have to do to account for clustering in the C-index calculations or is it sufficient that I used the predicated survival values from a cluster adjusted proc phreg to then calculate the C-index? I should note that the amount of clustering in my data is probably not significant - there are slightly less than 10% of deliveries are a second or third delivery for the mother.
Dear Herzog,
You can see the link below. It's http://gaodoris.blogspot.com/2012/10/5-ways-to-estimate-concordance-index.html
It's a practical method with R
Best regards
Thanks so much for your reply! I found that site in my search and it's very informative! Unfortunately, it's not clear to me the following:
Does the C-index ITSELF (or any measure of concordance on survival data) need to ALSO be adjusted for clustering. I think the answer is no, after much searching, and that is is sufficient to adjust your model and use the results of that model to calculate the C-index. But I'm not sure.
Hi Catherine,
Sounds like you are looking for something like a concordance statistic in the framework of shared frailty. Here is a description of a strategy along those lines using the R package "frailtypack."
http://virginierondeau.com/BiostatisticalConsulting/Publications_files/Mauguen2013StatMed_Concor.pdf
Given that the majority of your observations appear to be singletons, though, I would be less concerned about the clustering effect. If reasonable to ignore this, Frank Harrell's RMS package, as mentioned above, has an easy to implement "validate" function that provides a concordance statistic for time to event data which is adjusted for model selection and estimation optimism (getting at internal validation).
If you use Stata, you get the results immediately after just enter the .estat concordance command after performing cox regression
- Badges
- Science topic
- Similar topics
- Medicine
- Epidemiology