Skip to content

Second-order effect in Parametric Survival Analysis

4 messages · ryusuke, David Winsemius

#
Hi experts,

http://r.789695.n4.nabble.com/file/n4034318/Parametric_survival_analysis_2nd-order_efffect.JPG
Parametric_survival_analysis_2nd-order_efffect.JPG 
As we know a normal survival regression is the equation (1)
Well, I'ld like to modify it to be 2nd-order interaction model as shown in
equation(2)

Question:
Assume a and z is two covariates.
x = dummy variable (1 or 0)
z = factors (peoples' name)
fit <- survreg(Surv(time,censor)~x*z, data=sample, dist="exponential")

I tried to apply survreg(), while I have few questions:
1) If */survreg(Surv(time,censor)~x*z, data=sample, dist="exponential")/*
correct?
2) If the baseline hazard is the value excluded both x and z effects?
3) How can I get the value and plot the hazard with only x effect (but
exclude effect z)

Thanks


Best,
Ryusuke


--
View this message in context: http://r.789695.n4.nabble.com/Second-order-effect-in-Parametric-Survival-Analysis-tp4034318p4034318.html
Sent from the R help mailing list archive at Nabble.com.
#
On Nov 12, 2011, at 7:37 AM, ryusuke wrote:

            
The formula interface for R would expand x*z to x + z + x:z (Which is  
not the formula in your Nabble-provided-jpg, but from your later   
questions is probably what you want anyway.)
Maybe. You won't be  "excluding" them so much as holding their values  
jointly at zero, which may or may not be the same thing.
You will never be able to do so. If you have an interacting variable  
in a model, there will always be an effect of that covariate on  
predictions associated with any covariate with which it is  
interacting. You should be able to display or plot the ""x- 
effects" (note the plural) that are estimated for chosen levels of z,  
however. To accomplish that you should construct an appropriate  
data.frame and offer it as the newdata argument to predict(fit) ....  
just as you would do with any properly constructed R/S regression  
package.

There is a worked example on this posting from the Master:

http://finzi.psych.upenn.edu/Rhelp10/2010-May/240458.html
#
Thank you Dr. David.

I try to summarize it.
Assumes x and z are two covariates:
x = dummy variable (1 or 0)
z = factors (people name)

x*z = x + z + x*z
therefore this is not a 2nd-order interactions, it should be (for an
exponential survival regression):-
h(t|(X=x,Z=z)) = exp(Beta0 + XZBeta1)

#---------------------------------------------------

I believe there is no 2nd-order interactions survival regression as I
searched over www.rseek.org. While I tried to read through the codes of
survreg(), I stuck (cannot understand) at survreg6.c

survreg6.c apply C Language which involves Cholesky decomposition
multi-matrix (first-order interactions) calculation.
1) chinv2.c
2) cholesky3.c
3) chsolve2.c (only solve the equations of first-order interactions)

If someone gives some idea or suggestion on these?
Thank you.


Best,
Ryusuke


--
View this message in context: http://r.789695.n4.nabble.com/Second-order-effect-in-Parametric-Survival-Analysis-tp4034318p4036005.html
Sent from the R help mailing list archive at Nabble.com.
#
On Nov 13, 2011, at 12:51 AM, ryusuke wrote:

            
Actually I said = x + z + x:z

And interaction formula of a two level dummy with a multi-level factor  
would produce and intercept (which would be for the first person's  
name), a coefficient for each of other names at level zero, a dummy  
coefficient (for the first person), and interaction coefficients of  
each person at the 1-level.
If Beta1 is not a vector in this instance, with a distinct value for  
each(x,z) pairing, then I am unable to make sense out of that model.  
The questin remains however whether you are also expecting Beta0 to  
also be distinct for each specific combination of covariates.
That level of implementation should be addressed to a person with  
higher levels of knowledge: Therneau or Lumley are the two names that  
immediately come to mind.
David Winsemius, MD
West Hartford, CT