I haven't checked the web site, but several issues always spring to mind
when looking at "adjusted" outcomes.
1. If you average a covariate, what does it mean to be 0.52 male or 0.13
black?
2. Is the average over the estimation sample, a super-sample, or an outside
sample? What is the variance associated with the average?
3. Will a non-linear model produce adjusted estimates outside the range of
un-adjusted estimates? Does this make any sense?
I recommend (as usual) Graubard. B. and Korn, E. "Predictive Margins With
Survey Data" Biometrics, June 1999
Bryan Sayer
Statistician, SSS Inc.
-----Original Message-----
From: [email protected]
To: [email protected]
Cc: [email protected]
Sent: 7/10/03 9:57 AM
Subject: st: Adjusted survival question for Mario Cleves or Bill Gould
Stata Tech Support suggested this as the appropriate venue for posing
this
question:
Below is a web site and the abstract about a comparison of methods for
plotting
adjusted survival. I would be very curious as to what you think of this
method, since adjustfor() in sts graph I think is doing the "average
covariate
method". Is there any thought of implementing the alternate approach in
Stata?
(The web site gives an .ado file, but it is a bit clunky)
Hebe Quinton
Clinical Research
Dartmouth Medical School
(603) 650-7710
=========
http://www.ucalgary.ca/~hquan/adjsurv.html
PLOTTING ADJUSTED SURVIVAL CURVES FROM PROPORTIONAL HAZARDS MODELS: A
COMPARISON OF TWO METHODS.
WA Ghali, H Quan, R Brant, CM Norris, G van Melle, ML Knudtson, for the
APPROACH (Alberta Provincial Program for Outcome Assessment in Coronary
Heart
Disease) Investigators.
University of Calgary, Calgary, AB, Canada
APPROACH is a large inception cohort study which captures all patients
undergoing cardiac catheterization in Alberta, Canada. We used data on
11,804
patients from this study to examine the two year survival experience of
diabetics vs. non-diabetics, while controlling for covariates such as
left
ventricular ejection fraction, coronary anatomy, sociodemographic
variables,
and comorbidities. Here, we present a comparison of two methods for
calculating
covariate-adjusted survival curves.
The most commonly-used method for generating such curves is the "average
covariate method", in which the average values of covariates of interest
are
entered into a proportional hazards regression equation to generate
adjusted
survival estimates. We used data from the APPROACH study to compare the
survival curves generated by the "average covariate method" with those
generated by a newer method - the "corrected group prognosis method" -
in which
a survival curve for each level of covariates is calculated, after which
an
average survival curve is calculated as a weighted average of the
individual
survival curves. The resulting adjusted survival curves are shown below,
along
with the corresponding unadjusted survival curves for diabetics and
non-diabetics.
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
