Skip to content
Back to formatted view

Raw Message

Message-ID: <47CBC1D5.9020004@wittongilbert.free-online.co.uk>
Date: 2008-03-03T09:16:05Z
From: Calum
Subject: Problem plotting curve on survival curve (something silly?)
In-Reply-To: <47CB2F15.1070606@biostat.ku.dk>

> Calum wrote:
>
>> All of that  is very nice so far.  The  I followed bits and pieces of 
>> other peoples posts in the past to plot on a weibull regression...
>>
>>  > my_curve.Plac <- survreg( Surv(Survival, Censored==0)~ 
>> TreatmentGroup, subset=TreatmentGroup=="Placebo", data=TestData, 
>> dist='weibull')

Peter wrote:
> I'd take a hard look at the pweibull(...)  bit. Is "scale" really what 
> you want it to be? If coef(my_curve.Pred) is not a scalar, then it gets 
> recycled, which could easily cause oscillations.

Aha - found it.  Its from my attempt to subgroup the data before I 
discovered subset...  I now have a curve instead of an oscillation. Now 
possibly that curve is wrong... will need to do more reading!

For reference the line above should read:

   > my_curve.Plac <- survreg( Surv(Survival, Censored==0)~
  1, subset=TreatmentGroup=="Placebo", data=TestData,
  dist='weibull')

>> Also is it possible to get an R-squared type value for the fit of this 
>> curve from someplace?
>>
>> Finally (three questions in one!) the first two censored data points 
>> (1 in each group) are actually lost to follow-ups.  Should they be 
>> marked differently from censored?
>>   
> Customarily they are not. (I'm sure it is possible to speculate at 
> length about it, though.)
> 
Going off topic a bit - but did you mean customarily they are not 
censored or customarily they are not handled differently from censored!


> Nothing spectacularly incompetent this far... (I'm not happy with R^2 
> measures outside of linear models, or even within linear models, but 
> several well-reputed people do find them useful, so who am I to bicker?)

I'm not competent to argue.  But are you suggesting there is a better 
way to assess fit of the line to the data?  Thats what I want - Not 
being a statistician I'm not fussed how its done.  But If I'm going to 
extrapolate a line I'd like to know its a reasonable fit first (is that 
purely by eye?)  There is a p value reported by survreg but no idea how 
to interpret it ;-)