Skip to content
Prev 131253 / 398502 Next

weighted Cox proportional hazards regression

On Tue, 4 Dec 2007, Scott Bartell wrote:

            
You can get the answer you want with

coxph(Surv(pseudotime, cc)~x+strata(riskset)+offset(log(wt)), data=d2, robust=TRUE)

which is how countermatching usually seems to be done, and is what the 
original paper by Langholz & Borgan recommends.

I think it's right that weight=wt doesn't do the same thing.  The weights 
are not simple inverse-probability sampling weights, because the sampling 
units in this design are pairs, not individuals.  If we assume the 
distribution of x is the same across risk sets (which looks approximately 
true in your data) then the sampling weight for a pair is proportional to 
the number of eligible controls: ie, just your control weight.  Using 
these as weights for the pairs in the weight= argument I get 0.585 as the 
hazard ratio, reasonably close to your 0.63 given that this is a different 
estimator and given the assumptions.

 	-thomas
Thomas Lumley			Assoc. Professor, Biostatistics
tlumley at u.washington.edu	University of Washington, Seattle