rpietro
commented
12 years ago Tony, before you break it into weekly chunks, which is what you do when the variable keeps changing over time, i would try to break the tie. the mechanics goes like this:
- find the median time and split it there using survsplit
- run the model with the split time and the non-split time - the estimates should be similar, but the proportionality assumption should hold now.
see 5.1.2 under http://goo.gl/PDiX3
all that being said, I was looking at Therneau's book (this is considered the bible for survival), specially his chapter on how to deal with ties (see attached ties.pdf), and one of the things he says is that if the plot shows a horizontal line one could ignore the test statistic (the p value). when i looked at the age plot i was initially surprised since the line seems pretty horizontal despite the p value. is that the same pattern across the board? if so, given the sample size you might decide to simply ignore the ph assumption test and go with your first model
On Sat, Sep 29, 2012 at 10:22 PM, acastleberry notifications@github.comwrote:
I think with our Cox model we will probably have to build interactions between covariates and time to accommodate non-proportional hazards. John Fox' "Appendix to An R and S-PLUS Companion to Applied Regression" explains it well, but I don't quite follow how the start and stop times are derived. Here is the example code from the text to address PHA violation:
Call: coxph(formula = Surv(start, stop, arrest.time) ~ fin + age + age:stop + prio, data = Rossi.2)
In the example from the text is sounds like they had weekly blocks of data and therefore had start and stop times based on that, but I'm not sure how to translate that to our survival analysis? The other text ( http://goo.gl/PDiX3) referred to partitioning the survival data in order to derive start and stop times, but I couldn't follow the mechanics of that. Is partitioning the way this is typically done?
— Reply to this email directly or view it on GitHubhttps://github.com/rpietro/airwayDehiscence/issues/15.
I think with our Cox model we will probably have to build interactions between covariates and time to accommodate non-proportional hazards. John Fox' "Appendix to An R and S-PLUS Companion to Applied Regression" explains it well, but I don't quite follow how the start and stop times are derived. Here is the example code from the text to address PHA violation:
Call: coxph(formula = Surv(start, stop, arrest.time) ~ fin + age + age:stop + prio, data = Rossi.2)
In the example from the text is sounds like they had weekly blocks of data and therefore had start and stop times based on that, but I'm not sure how to translate that to our survival analysis? The other text (http://goo.gl/PDiX3) referred to partitioning the survival data in order to derive start and stop times, but I couldn't follow the mechanics of that. Is partitioning the way this is typically done?