Expected number of events, Andersen-Gill model fit via coxph in package survival

I am interested in producing the expected number of events, in a
recurring events setting. I am using the Andersen-Gill model, as fit
by the function "coxph" in the package "survival."

I need to produce expected numbers of events for a cohort,
cumulatively, at several fixed times. My ultimate goal is: To fit an
AG model to a reference sample, then use that fitted model to generate
expected numbers of events for a new cohort; then, comparing the
expected vs. the observed numbers of events would give us some idea of
whether the new cohort differs from the reference one.

From my reading of the documentation and the text by Therneau and

Grambsch, it seems that the function "survexp" is what I need. But
using it I am not able to obtain expected numbers of events that match
reasonably well the observed numbers *even for the same reference
population.* So, I think I am misunderstanding something quite badly.

You've hit a common confusion.  Observed versus expected events computations are done on 
a cumulative hazard scale H, not the surivival scale S; S = exp(-H).  Relating this back 
to simple Poisson models H(t) would be the expected number of events by time t and S(t) 
the probability of "no events before time t".  G. Berry (Biometrics 1983) has a classic 
ane readable article on this (especially if you ignore the proofs).

   Using your example:

 > cphfit <- coxph(Surv(start,stop,event)~rx+number+size+cluster(id),data=bladder2)
 > zz <- predict(cphfit, type='expected')
 > c(sum(zz), sum(bladder2$event))
[1] 112 112

 > tdata <- bladder2[1:10]   #new data set (lazy way)
 > predict(cphfit, type='expected', newdata=tdata)
  [1] 0.0324089 0.3226540 0.4213402 1.0560768 0.6702130 0.2163531 0.6490665
  [8] 0.8864808 0.2932915 0.5190647


  You can also do this using survexp and the cohort=FALSE argument, which would return 
S(t) for each subject and we would then use -log(result) to get H.  This is how it was 
done when I wrote the book, but the newer predict function is easier.

Terry Therneau