Hi all,
I have a binary variable where I am comparing children with no congenital anomaly to those with certain types of anomaly and trying to model a survival analysis with death as the event. The variable has failed the proportional hazards test. I have tested interactions and have one. I am modelling in weeks over 5 years from birth. The log rank test is significant p<0.05 for the variable. I've done a test for a time varying covariate and tested against a model without this. I get improvement with the likelihood ratio test between models with the time varying covariate. The model line with a certain congenital anomalies is very steep at the beginning and then gets very flat after 1 year months (year 1 from 100-93%, years 1 to 5 from 93-90% survival). I can't use stsplit as the sample size is too large. When I use time (_t) as the time varying covariate the proportions are out by 10% by the end of the model when comparing actual proportions that have died to the survival analysis model. (ln(_t)) as the time varying covariate gives a better fit but I'm unable to test if this is a good fit.
Questions which I'd really like some help with:
1. What tests should I apply to (ln(_t)) time varying covariate to check for goodness of fit in Stata (I could convert to R)?
2. Is it worth trying (sqrt(_t)), are there any other suggestions for transformations I could use for a model of this shape?
3. Stratifying on the variable will not give me the comparison results I need in this case, is it possible to partition the time axis and report results separately? Are there any interpretation issues with this?
4. I've read about accelerated failure time models - would this be appropriate in this case?
5. I have looked into parametric modelling using weibull, loglogistic, inverse exponential - is it better to move to one of these distributions than to continue trying to fit the cox model?
Any help would be greatly appreciated!
many thanks,
Annette
You may leave the list at any time by sending the command
SIGNOFF allstat
to [log in to unmask], leaving the subject line blank.
|