Thank you to all who made very helpful suggestions to get started with R.
Duncan Murdoch raised an excellent question, asking about my background and
reason for using R. I'm an epidemiologist, applying the marginal structural
models approach (inverse probability of treatment weights) in a Cox
proportional hazards analysis.
The statistical program which I had been using does not have functions for
doing a weighted Cox analysis. Any advice pointing me in the direction of
resources to help probability-of-treatment-weighted Cox proportional
hazards would be most appreciated.
Best wishes,
Michael
____________________________________
Michael McCulloch
Pine Street Clinic
Pine Street Foundation
124 Pine Street, San Anselmo, CA 94960-2674
tel 415.407.1357
fax 415.485.1065
email: mm at pinest.org
web: www.pinest.org
www.pinestreetfoundation.org
www.medepi.net/meta
using R for survival analysis
2 messages · Michael McCulloch, Thomas Lumley
On Tue, 5 Dec 2006, Michael McCulloch wrote:
Thank you to all who made very helpful suggestions to get started with R. Duncan Murdoch raised an excellent question, asking about my background and reason for using R. I'm an epidemiologist, applying the marginal structural models approach (inverse probability of treatment weights) in a Cox proportional hazards analysis. The statistical program which I had been using does not have functions for doing a weighted Cox analysis. Any advice pointing me in the direction of resources to help probability-of-treatment-weighted Cox proportional hazards would be most appreciated.
You can do IPW estimation with the coxph() function in the 'survival' package as long as you use the `robust=TRUE' option for standard errors. The weights can be supplied as weights to coxph(). You need to fit logistic regression models for treatment, with glm(), then use predict() to get probabilities of treatment (yes/no) then convert these into probability of observed treatment, then take the reciprocal to get the weights. How messy this all is depends on the data structure -- how regular the observations of 'treatment' are, for example -- since this determines how many different logistic models for treatment you will need. -thomas Thomas Lumley Assoc. Professor, Biostatistics tlumley at u.washington.edu University of Washington, Seattle