[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

st: clustering in proportional hazards models with stata/mp 10.0

From   "Daniel O. Koralek" <[email protected]>
To   [email protected]
Subject   st: clustering in proportional hazards models with stata/mp 10.0
Date   Wed, 5 Sep 2007 18:35:48 -0400

Please excuse this if it is a double post, but I don't believe my first attempt made it to the listserv...


Please forgive me, but I'm a bit of a novice about how to treat a variable in some survival analysis. I have an observational study where individuals were recruited at 1 of 10 centers. I would like to account for center because it may influence my other covariates (e.g. different foods eaten in different regions may influence nutrients, which are my main exposures of interest). In past analyses, I simply controlled for center as a categorical covariate in my cox models. But I'm thinking that it would be better to do something a little better. Due to confidentiality with respect to the data, I can't present the numbers that I get here.

Using the vce option in stcox (e.g. "vce(cluster center)" ) does what I would expect and reduces the variance of my estimates, but doesn't change the estimate themselves. Using the strata option (e.g. "strata(center)") changes the estimates (am I to assume that this is similar to allowing for random intercepts in a linear model, in this case modeling separate baseline hazards for each level of the strata variable? or I could just be thinking about things in that analogy incorrectly). Now, one could use both strata and vce in the same model, and I get the types of changes I think make sense, the same point estimates as using the strata option with smaller variance that i would expect with the vce option. Is this combination kosher? and is it a good way to think about handling this type of variable?

examples of code would be:

just the categorical nutrient variable:
xi: stcox i.cat_nut, efron

using the strata option:
xi: stcox i.cat_nut, efron strata(center)

using the clustering option:
xi: stcox i.cat_nut, efron vce(cluster center)

using both:
xi: stcox i.cat_nut, efron vce(cluster center) strata(center)

or the way I would have done it without thinking, would have been:
xi: stcox i.cat_nut, efron

Any suggestions on how best to handle this type of covariate would be appreciated.



Daniel O. Koralek
Department of Epidemiology/Lineberger Comprehensive Cancer Center
The University of North Carolina at Chapel Hill
Chapel Hill, NC 27599-7435
* For searches and help try:

© Copyright 1996–2024 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index