Thanks Joseph,
Yes the problem exists in most of my models. The groupings are
time-varying as the geographic areas were adjacent, and animals were
free to come and go as they pleased.
I'm wondering if a solution would be to treat each group as a strata?
Thus each group would have its own baseline hazard function.
HR
-----Original Message-----
From: Joseph Coveney [mailto:[email protected]]
Sent: Thursday, July 31, 2008 6:24 PM
To: Statalist
Cc: Hugh Robinson
Subject: Re: st: Cox Model Collinearity
Hugh Robinson wrote:
I am trying to model wildlife survival data (multiple encounter records
for 117 individual cougars over 9 years). I have divided the data into
several "groupings" based on varying hunting pressure and geographic
location. I wish to use Cox modeling and post estimation AIC (estat ic)
to test which grouping best fits the mortality patterns observed. For
instance, all areas were hunted heavily for the first 3 years (group 1)
at which point one area was protected (group 2) while hunting continued
in the remainder of the study area (group 3). A second model would test
if dividing the study animals into 6 groups (i.e. 2 geographic areas, 3
time periods) provides a better fit. Each group is coded as a string
variable.
The outputs I receive from STATA suggest that several groups are
collinear resulting in no values of Standard Error and erroneous hazard
ratios. For instance, in the above model group 3 and group 1 are
collinear. With group 1 coded as the dummy variable, STATA provides a
hazard ratio of 0.99 and no standard error, z, or p-value. I'm not sure
if I can trust the AIC values I receive from post estimation.
Can anyone suggest how I should approach this collinearity issue?
------------------------------------------------------------------------
--------
Grouping two geographic areas together at Time 1 as Group 1, and then
separately at Time 2 as Groups 2 and 3 wouldn't give rise to
collinearity by
itself. If the problem is caused by following cougars in the same
geographic area through both times, then you would be seeing the same
problem in the second model. I take it that you are seeing the problem
with
both models--is that correct? If so, have you looked into what Stata
has
for time-varying covariates (predictors) in Cox proportional-hazard
models?
I'm not sure how helpful it will be in your circumstances, but if you
haven't already seen what's available, then type
findit cox time varying
in Stata's command window and follow the hyperlinks.
The commands there might allow for a better parameterization of your
model.
Joseph Coveney
*
* For searches and help try:
* http://www.stata.com/help.cgi?search
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
