[R] ... predict.coxph

Mike Marchywka marchywka at hotmail.com
Mon Nov 15 16:10:45 CET 2010

> From: therneau at mayo.edu
> To: james.whanger at gmail.com
> Date: Mon, 15 Nov 2010 08:43:04 -0600
> CC: r-help at r-project.org
> Subject: Re: [R] ... predict.coxph
> >If you are looking at radioactive decay maybe but how often do
> >you actually see exponential KM curves in real life?
> Exponential curves are rare. But proportional hazards does not imply
> exponential.

Well, I'm being a bit extreme but of course the point would
be as you being deviating from that the fit could arguably
become less connected to anything of causal relevance. 
You've probably seen exchanges I've had with people wondering
why their linear regression doesn't work over arbitrary 
data blobs, how much more so here. A time constant that you
can relate to tunnelling probability has some relevance. 

> > A trial design
> could in fact try to get all the control sample to "event" at the same
> time if enough was known about prognostic factors and natural trajectory
> You are a dreamer. We know very little about even the diseases we know
> best. The life insurers are in no danger yet from accurate predictions
> by the medical community.

OBviously the quesion is a bit hypotheticl, but thinking ahead
I wonder what kind of curves you are likely to see and how they 
relate to each other. I just thought if you defended the model,
and I admit to only skimming prior conversation, you may
be willing to volunteer additional summary comments.
I guess you could take something like a sigmoidal survival
curve and various limits of that. Certainly this won't happen
tomorrow but presumably the goal here is to reduce random/noise/variance
 and in any case use statistics to estimate
parameters that can relate to something physically measurable ( blood
parameters or something). In any case, what kinds of things could
a putative treament do to a curve like this that would not be
reasoably described by coxph ? Is this not something that you 
ever expect to care about? I guess my point is that the control
and treatment curves needs to have some relationships to make
ph parameters fit to anything relating to somehting causal. 

I'd also note that insurers typically need a prognosis when there is
no detectable evidence of a disease but trial criteria may have
some range of disease-range parameters that will allow enrollment.
Picking out small deviations from normal to first identify the
diseaes that will get you and then the time to death is of course
much more difficult than plotting a trajectory when someone has
clinical manifestations of a speicific diease. 

> Terry T

More information about the R-help mailing list