[R] counting process form of a cox model (cluster(id))

[Terry Therneau]

Zoe writes: 
My question is quick.  I am looking at 1 event (death), and repeated 
measurements (the time dependent covariate 'lqol') are frequently taken on a 
subject, so I assume that measurements on the same subject will be correlated. 

The answer is: no, it's not a problem
When the time intervals for a subject are disjoint, e.g, 0-10, 10-49, 49-127,
etc, like they will be on this data, the mulitple lines are just a computational
trick.  Any given term in the likelihood will select the right line of 
data for each person, but only one line.  
   Since the multiple rows of data for a person never appear together, it
does not matter if they are correlated or not.  The set of lines that are
chosen for the likelihood have only 1 (or zero) appearances for each person,
hence are an independent set of observations.  So you don't need the robust
variance.
   However, if you allow time travel, e.g. a person returns to time zero after
an event, that is another kettle of fish.  You then have two copies of the
same person at the same party at the same time, and they can interact.  You 
will need a robust variance, but also want to think hard about whether the
model itself makes any sense.
   If there are multiple events per person then one needs the sandwich variance,
but for a somewhat different reason. 

Terry Therneau