R help - May 2011 - Longitudinal data with non-randomized subjects

Dear List,
I have a theoretical question related to epidemiological data analysis: 
If the treatment status (tx = 0,1) changes over time for the patients in a
non-randomized cohort, is there a way to estimate the treatment effect?
(i.e., after joining the study, some patients may have to wait for a period of
time before receiving the treatment, i.e., the situation of patient with id == 2
for the following data)
Data format is like the stanford heart transplant data (Therneau et al 2000,
p69), but the patients were not randomized in selection and the covariate
balance is not achieved:
id	time	censor	tx	x1	x21	(0,10]	1	0	x11	x122	(0,8 ]	0	0	x21	x222	(9,19]	1	1	x21
x223	(0,13]	0	1	x31	x32
Is counting process form of a Cox model (coxph with start, stop, censoring
status ~ tx + x1 + x2 covariates) sufficient?
Is it possible to implement the propensity score methodology (Rosenbaum et al,
1983) in such situations?
Any ideas/suggestions would be higly appreciated.
Thanks,
Ehsan 		 	   		  
	[[alternative HTML version deleted]]

Apology for reposting, but the format of earlier message got
distorted; hopefully this time it will be readable:


From: wildscop at hotmail.com
To: r-help at r-project.org
Subject: Longitudinal data with non-randomized subjects
Date: Sun, 1 May 2011 00:34:08 -0700

Dear List,

I have a theoretical question related to epidemiological data analysis:

If the treatment status (tx = 0,1) changes over time for the patients
in a non-randomized cohort, is there a way to estimate the treatment
effect?

(i.e., after joining the study, some patients may have to wait for a
period of time before receiving the treatment, i.e., the situation of
patient with id == 2 for the following data)

Data format is like the stanford heart transplant data (Therneau et al
2000, p69), but the patients were not randomized in selection and the
covariate balance is not achieved:

id	time	censor	tx	x1	x2
1	(0,10]	1	0	x11	x12
2	(0,8 ]	0	0	x21	x22
2	(9,19]	1	1	x21	x22
3	(0,13]	0	1	x31	x32

Is counting process form of a Cox model (coxph with start, stop,
censoring status ~ tx + x1 + x2 covariates) sufficient?

Is it possible to implement the propensity score methodology
(Rosenbaum et al, 1983) in such situations?

Any ideas/suggestions would be higly appreciated.

Thanks,

Ehsan