R help - Mar 2008 - Problem plotting curve on survival curve

Calum had a long question about drawing survival curves after fitting a Weibull 
model, using pweibull, which I have not reproduced.

It is easier to get survival curves using the predict function.  Here is a
simple example:
> library(survival)
> tfit <- survreg(Surv(time, status) ~ factor(ph.ecog), data=lung)
> table(lung$ph.ecog)
   0    1    2    3 <NA> 
  63  113   50    1    1 
> tdata <- data.frame(ph.ecog=factor(0:3))
> qpred <- predict(tfit, newdata= tdata, type='quantile',
p=1:99/100)
> matplot(t(qpred), 99:1/100, type='l')
  The result of predict is a matrix with one row per group and one column per 
quantile.  The final plot uses "99:1" so as to show 1-F(t) = S(t)
rather than F.
Don't ask for the 1.0 quantile BTW -- it is infinity and I doubt you want
the
plot to stretch out that far.  The 0.0 quantile can also have issues due to the 
implicit log transform used in many distributions.  
   If I had not used the newdata argument, we would get 227 rows in the result, 
one for each subject.  That is, 63 copies of the ph.ecog==0 curve, 113 of the 
ph.ecog==1 curve, ...  The above fit assumed a common shape for the 4 groups, 
you can add a "+ strata(ph.ecog)" term to have a separate scale for
each group;
this would give the same curves as 4 separate fits to the subgroups.
  
  There are several advantages to using the predict function.  The first is that
the code does not need to change if you decide to use a different distribution.
The second is that you can add the "se.fit=T" argument to get
confidence bounds
for the curves.  (A couple more lines for your matplot call of course).

	Terry Therneau
	Mayo Clinic

Terry Therneau wrote:
> 
> It is easier to get survival curves using the predict function.  Here is a
> simple example:
>> tfit <- survreg(Surv(time, status) ~ factor(ph.ecog), data=lung)
>> tdata <- data.frame(ph.ecog=factor(0:3))
>> qpred <- predict(tfit, newdata= tdata, type='quantile',
p=1:99/100)
>> matplot(t(qpred), 99:1/100, type='l')
> 
Many thanks - that worked at treat... (One day I might work out what it 
does - for now I'm happy it does it!) In terms of when I write up what I 
did is this still a weibull regression?  help(predict.survreg) just 
calls it a quantile...  (Sorry that may be dumb question ;-) )
> The above fit assumed a common shape for the 4 groups, 
> you can add a "+ strata(ph.ecog)" term to have a separate scale
for each group;
> this would give the same curves as 4 separate fits to the subgroups.
Any thoughts on which is scientifically more valid?  I'd have thoughts 4 
separate shapes?  Certainly if I'm modeling drugs - its surely possible 
that a new drug might change the course of disease and therefore the 
shape of the curve altogether?

Brings me back to my extra question - is there any way to determine 
quality of the fit for this (like an R^2 value for a linear regression). 
  That might answer if a strata approach is needed.